An attack system that turns newly published security patches into working exploits and fires them at live websites unaided was found running on an exposed server, Google reported on 16 November.
The system, called exploit_agent, was wired into several commercial and locally hosted language models and compared the code of popular web applications before and after each patch. When it identified a likely flaw it drafted an exploit and ran it. The same server held stolen data and scripts for guessing passwords and planting web shells, small programs that give an intruder remote control of a web server. The victims were mainly French media, entertainment and education websites. The operator is unidentified, and its configuration file carried a Russian-language comment. Google's Threat Intelligence Group described exploit_agent as the first fully automated pipeline it has observed in use, two months after a quarterly report that recorded only partly automated attacks.
Three days earlier the FBI had issued an alert, FLASH-20261113-001, on ARTEX, an open-source AI tool written for authorised penetration testing that criminals and state-aligned groups have turned on banks, a software supply chain and organisations in 10 sectors. Some intrusions went from initial targeting to privileged access in under three hours, according to the alert. Over the past 19 months, authorities in Norway, Poland and France have blamed state or state-aligned actors for cyber sabotage of a dam, energy plants and supermarket refrigeration.
Published findings on the dam, energy and refrigeration attacks describe human operators working through exposed interfaces, weak passwords and custom malware. exploit_agent is the single documented system that ran a full attack chain unsupervised, and a human built it and chose its targets. Akamai, the internet infrastructure company whose research the FBI cites, assessed that AI changes the speed of attack while the techniques stay the same. Google's researchers wrote that the agent-driven tools they have seen leave noisy forensic trails, which for now makes them easier to detect.
A patch becomes a weapon in days
Software makers publish a patch to close a flaw, and the patch shows anyone who compares the old and new code where the flaw was. Defenders have relied on the days or weeks attackers needed to turn that comparison into working attack code. Google's researchers wrote that attackers using language models may now find it more efficient to attack flaws already disclosed and patched, known as n-days, than to hunt for unknown ones, because patches arrive in a steady and predictable supply.
In late August two chained flaws in PaperCut print management servers were exploited before a fix existed, and security researchers, including at Huntress, used language models to build working exploits within a day of disclosure. A separate campaign tracked by GreyNoise, which monitors internet-wide scanning, began four days after a disclosure and hit hundreds of organisations, using exploits that Google's report says were written with language models. In one victim's network the intruder reached domain administrator rights, the master keys to a corporate Windows network, within minutes.
Mandiant, Google's incident response arm, described what it considers the first confirmed zero-day vulnerability developed with AI in its September report. In May a criminal used an AI model to help discover and weaponise a way around the two-factor login of an open-source web administration tool, intending to exploit it at scale. The same report found that exploiting software flaws was the most common way into victims' networks for the sixth year running.
ARTEX, from GitHub to 10 sectors
A developer using the name Autumn-27 published ARTEX on GitHub in July 2026. Malicious use began by August. On 8 October the developer deleted the repository and wrote that the project would be frozen and converted to closed source. Copies still circulate through forks, archives and a paid edition, ARTEX PRO, and researchers counted more than 1,600 ARTEX control consoles exposed on the internet in early November.
An ARTEX operator states an objective. A planning agent breaks it into tasks, and worker agents run scanners, password-guessing tools and scripts while a shared record tracks what has been found. Cisco Talos, the company's threat intelligence unit, wrote that ARTEX needs little input from its operator once a task is under way. Attackers supply the language model themselves, and the FBI found they mostly chose Chinese-origin models, favouring DeepSeek for its capability, low cost and lack of observable guardrails.
Early in October an ARTEX operator took the personal information of about 67,000 people from South Korean financial institutions. CrowdStrike and Zscaler each assessed with moderate confidence that the operator was a financially motivated Chinese speaker. Storm-0940, a group Microsoft assesses as aligned with the Chinese state, used stolen access to ConnectWise Cloud Backup to try to sign in to more than 2,250 downstream Microsoft 365 customers. The FBI says the group used ARTEX to speed up scanning, password spraying and abuse of applications already granted access to customers' cloud accounts.
The FBI lists financial services, government, technology, defence, semiconductors, energy, healthcare, manufacturing, telecommunications and law among the affected sectors. Citing Akamai, the alert says ARTEX works best against systems organisations watch least: supplier portals, mobile app back ends, staff tools, legacy services and applications built by individual departments. The FBI wrote that ARTEX automates intrusions "at machine speed, from reconnaissance through data exfiltration" and lets actors with minimal technical expertise attack many targets at once.
Talos found ARTEX used mainly by financially motivated criminals and by actors linked to China. Google and Mandiant have seen it and two similar open-source tools, PentAGI and CAIRN, in incident response cases since late August. One actor used CAIRN during an intrusion against the U.S. aerospace sector. Another ran a customised PentAGI from late August to early October to scan organisations in Asia. Google's researchers recorded ARTEX checking permissions across Amazon Web Services security groups in 27 events in 23 seconds, and recorded its mistakes as well.
HexStrike AI, released in August 2025, connects language models to more than 150 security tools. Within hours of Citrix disclosing flaws in its NetScaler gateways on 26 August 2025, criminals on underground forums discussed using HexStrike against them, Check Point Research reported. Some wrote that it cut the time to exploit from days to under 10 minutes, a claim that rests on the forum posts Check Point relayed.
How much is autonomous
A human penetration tester works in stages. Reconnaissance finds what an organisation has exposed to the internet. Vulnerability identification matches software versions against known flaws. Exploitation runs code that abuses a flaw to get in. The intruder then steals passwords and access tokens from the first machine and uses them to reach others, known as lateral movement, until holding administrator rights. In the AI tools, a language model decides which program to run next and reads the result. Repeating that loop thousands of times at machine pace compresses days of work into hours.
Anthropic gave the most detailed public account of such a loop in a state operation in November 2025. A group it designated GTG-1002 and assessed as Chinese state-sponsored manipulated the company's Claude Code assistant into acting as a team of penetration testing agents against about 30 technology companies, financial institutions, chemical manufacturers and government agencies in mid-September 2025. The AI carried out 80 to 90 per cent of tactical operations independently, Anthropic found, while human operators chose targets and approved each escalation. It also wrote that Claude "frequently overstated findings and occasionally fabricated data". Security researchers quoted by BleepingComputer doubted the autonomy figure.
Amazon's threat intelligence team documented a far less skilled operator. Between 11 January and 18 February 2026, a Russian-speaking, financially motivated individual or small group used at least two commercial AI services to compromise more than 600 FortiGate firewalls in more than 55 countries. The operator worked entirely through management ports exposed to the internet and weak passwords on devices protected by a single login factor. Credential databases and backup systems followed, a sequence that commonly precedes ransomware. CJ Moses, Amazon's chief information security officer, called it "an AI-powered assembly line for cybercrime".
Amazon's researchers found that the same actor's attempts to use known exploits failed repeatedly against better-maintained devices. In every case documented by Google, the FBI, Anthropic and Amazon, a human started the operation, chose or approved the targets and received the data or money.
Food, heat and water
On the night of 11 to 12 October, attackers entered refrigeration controllers in about 700 of France's 10,000 supermarkets through their web interfaces, set defrost cycles to run indefinitely and switched off temperature alarms. Display cabinets reached 50°C for about six hours, the Associated Press reported. Intermarché reported losses of tens of millions of euros in a day, and France's Cyber Crisis Coordination Centre counted at least 15 fires. Vincent Strubel, director general of ANSSI, the national cybersecurity agency, said the sites were reached through direct internet connections and weak passwords. Danfoss, maker of most of the controllers, said the stores "were running outdated firmware".
The French foreign ministry attributed the attack on 9 November to Shahid Kaveh, a unit of the Electronic and Cyber Centre of Iran's Islamic Revolutionary Guard Corps. Investigators cited shared infrastructure, traces of Farsi and IP addresses linked to devices in Iran. Iran's foreign ministry called the accusation "baseless and one-sided". The French government has kept its forensic evidence confidential, so the attribution rests on the word of the state and its agencies.
A second wave on 1 and 2 November changed refrigeration settings at stores in Belgium, Luxembourg, the Netherlands, Sweden, Ireland, Estonia, Finland, Portugal and Italy. Musgrave withdrew chilled and frozen goods from 13 of its 223 SuperValu stores in Ireland, and an ICA store in Sjöbo, in southern Sweden, found its freezers set to 50°C. Three cold-storage sites of the French blood service were manipulated, but separate monitoring prevented any loss of blood products. Two pro-Russian groups, NoName057(16) and Z-Pentest Alliance, claimed the wave on 4 November. DomainTools, a threat intelligence company, assessed the claim as propaganda unless the groups produce proof. The authors of the November wave remain unidentified.
On 29 and 30 December 2025, attackers inside the networks of more than 30 Polish wind and solar farms damaged the firmware of remote terminal units, the field devices that report to grid operators, and ran custom wiping malware. CERT Polska, the national response team, described the plan as partially automated. The farms lost remote control and contact with the distribution operator but kept generating. At a combined heat and power plant supplying almost half a million customers, the intruders had spent a long period stealing data and privileged accounts. When they triggered the wiper, the plant's endpoint security software blocked it.
CERT Polska found a high degree of infrastructure overlap with the cluster known as Berserk Bear or Static Tundra, and called the Polish attack the first destructive activity publicly attributed to that cluster. ESET, a Slovak security company, attributed the attack to Sandworm, a unit of Russian military intelligence. Prime Minister Donald Tusk said the attacks appeared to have been prepared by groups directly linked to Russian services.
On 7 April 2025, intruders took control of a dam at Bremanger in western Norway and opened a floodgate, releasing 500 litres of water a second for four hours before anyone noticed. The reservoir was well below flood capacity, which kept the release harmless. Beate Gangaas, head of the Norwegian Police Security Service, attributed the attack to pro-Russian cyber actors in August 2025. "Our Russian neighbour has become more dangerous," she said. The Russian embassy in Oslo called her statements unfounded and politically motivated.
Hospitals, airports, telecoms and banks
King's College Hospital NHS Foundation Trust confirmed in June 2025 that a patient who died unexpectedly during the 2024 ransomware attack on Synnovis, a pathology contractor, had waited a long time for a blood test result because of the attack. The delay was one of several factors in the death. The attack, claimed by the criminal group Qilin, forced the postponement of 10,152 outpatient appointments and 1,710 elective procedures at two London hospital trusts. On 19 February 2026, ransomware took down the medical records system of the University of Mississippi Medical Center and forced the closure of its clinics across the state.
Ransomware on Collins Aerospace's MUSE check-in software from 19 September 2025 forced Heathrow, Brussels, Berlin Brandenburg and Dublin airports onto manual processing, and Britain's National Crime Agency arrested a man in West Sussex five days later. An attack on Jaguar Land Rover halted its production for five weeks from 1 September 2025. The Cyber Monitoring Centre, an independent British body, put the UK cost at about £1.9 billion and the number of affected organisations above 5,000. Britain's National Cyber Security Centre recorded 204 nationally significant incidents in the year to August 2025, up from 89 the year before.
Salt Typhoon, a Chinese state-backed espionage campaign, had penetrated organisations in more than 80 countries, including more than 200 U.S. targets, by 27 August 2025. Agencies from 13 countries that day named three Chinese companies serving the People's Liberation Army and the Ministry of State Security. The intruders entered telecoms carriers through unpatched routers and gained the ability to track communications. The FBI said North Korea stole about US$1.5 billion from the exchange Bybit on 21 February 2025. Google's 16 November report says UNC6836, a suspected North Korean group, altered settings files in companies' code repositories so that each launch of Anthropic's Claude Code assistant downloaded a backdoor.
JPMorgan Chase traced the disruption to its online banking across the U.S. on 16 November to a software bug in its own hardware. The Federal Aviation Administration traced the nationwide ground stop of 13 November to old equipment that became unstable during maintenance.
Russia, China, Iran, North Korea and the ransomware groups
Polish and Norwegian authorities and ESET tie the energy and dam attacks to Russian state units or pro-Russian actors. NoName057(16), a pro-Russian group that Europol has described as mainly Russian-speaking sympathisers using automated tools, claims denial-of-service attacks on European governments and utilities.
Chinese state-linked groups dominate documented espionage: Salt Typhoon in telecoms, Storm-0940 in the ConnectWise supply chain and GTG-1002 in the campaign Anthropic exposed. North Korea's documented operations are theft and infiltration of technology companies. Ransomware caused the Synnovis, Mississippi and Collins Aerospace disruptions.
AI-built personas and account farms
A commercial contractor used Google's Gemini service to help build a platform designed to run more than 100,000 social media accounts from cloud phone farms, Google reported on 16 November. A separate platform aimed at political sentiment in Europe and the Middle East prompted its image generator for deliberate camera flaws so that synthetic photographs would pass as smartphone snapshots. Google disabled the accounts tied to the misuse of its service. Recruiter profiles that UNC6836, the suspected North Korean group, used to approach cryptocurrency companies in North America and Europe carried photographs bearing SynthID, the invisible watermark Google embeds in images its own models generate.
Grid failures in Iberia and Texas
Water treatment, telecommunications, card payments, hospitals and refrigerated food chains all draw on the electricity grid. The UK government's 2026 National Risk Register says a national power outage could set off a rapid domino effect across water, communications and transport. It asks infrastructure operators to test their plans against power and water failing at the same time. The register models a malicious cyberattack on the electricity system as one possible cause.
Its reasonable worst case for a total failure of Britain's transmission system assumes a winter collapse in which every home and business lacking a generator loses supply instantly. Restoring the network takes up to seven days in that scenario, and some services take months. The register anticipates loss of life and leaves it unquantified.
The power systems of continental Spain and Portugal collapsed at 12:33 CEST on 28 April 2025. An expert panel of ENTSO-E, the association of European grid operators, traced the total blackout in its final report of 20 March 2026 to overvoltage and several interacting failures of voltage control. Trains and metros stopped, petrol stations closed and hospitals switched to generators and suspended routine work, the Associated Press reported.
Three members of one family in Taboadela, in Galicia, died of carbon monoxide poisoning from a generator they were using to run an oxygen machine, Spanish emergency services said. Some regions of the peninsula waited up to 16 hours for power. The blackout had a technical cause, and the British register assumes the same chain of consequences for a cyberattack on a national grid.
Electricity and natural gas failed together in Texas in February 2021, when Winter Storm Uri froze gas supply to power stations and power cuts in turn stopped gas compressors. More than 4 million customers lost electricity at the peak, water systems lost pressure and millions of people were put under boil-water notices. Texas health officials put the death toll at 246.
Ukraine, Colonial Pipeline and NotPetya
On 23 December 2015, attackers who had spent about six months inside the networks of three Ukrainian electricity distributors cut power to about 225,000 customers in attacks launched within 30 minutes of one another, according to a NERC analysis. They also flooded the companies' call centres with thousands of calls to block customers reporting outages. Engineers restored supply within hours by operating circuit breakers by hand. The attack is attributed to Sandworm.
On 12 December 2023, Sandworm wiped much of the infrastructure of Kyivstar, Ukraine's largest mobile operator, and cut mobile and internet service for about 24 million subscribers. Ukraine's cyber security chief said the intruders had been inside the network since at least May 2023. The outage disrupted air raid sirens, banks, cash machines and card terminals that depended on Kyivstar connections, The Record reported. About 30 per cent of PrivatBank's cashless terminals stopped working, according to the Atlantic Council.
On 7 May 2021, ransomware hit the business computers of Colonial Pipeline, and the company shut its 5,500-mile pipeline, which carries 45 per cent of the fuel consumed on the U.S. East Coast. It closed the pipeline as a precaution, to keep the infection away from the systems that control fuel flow. Fuel ran short and prices rose across the region until the pipeline restarted on 12 May, six days later.
NotPetya, the wiper Sandworm hid in a Ukrainian accounting software update in June 2017, caused more than US$10 billion in damage worldwide, according to a White House assessment. The shipping company Maersk reinstalled 4,000 servers, 45,000 PCs and 2,500 applications in 10 days and moved about 80 per cent of its usual volume by hand while it rebuilt.
UNFI, a grocery distributor serving more than 30,000 locations in North America, halted deliveries after detecting intruders on 5 June 2025, and Whole Foods stores reported empty shelves. On 19 July 2024 a faulty CrowdStrike security update crashed about 8.5 million Windows devices and disrupted airlines, hospitals and emergency call services in several countries.
Medicare mortality and the San Diego diversions
Hannah Neprash, Claire McGlave and Sayeh Nikpay of the University of Minnesota linked records of hospital ransomware attacks to Medicare claims. Their study in the American Economic Journal: Economic Policy found that in-hospital mortality among patients already in hospital when an attack began rose by 34 to 38 per cent. Hospital volume fell by 17 to 24 per cent in the first week of an attack.
Ransomware shut Scripps Health's systems in San Diego for a month from 1 May 2021. Daily visits to two nearby emergency departments rose by 15.1 per cent and ambulance diversions and waiting times increased, according to a study in JAMA Network Open led by Christian Dameff of UC San Diego. The nearby hospitals treated 103 stroke patients in the four weeks of the attack, against 60 in the four weeks before. The authors wrote that such attacks should be treated as regional disasters.
Water controls and pre-positioned access
From 22 November 2023, attackers affiliated with Iran's Islamic Revolutionary Guard Corps, calling themselves CyberAv3ngers, took over Unitronics controllers at several U.S. water facilities. A joint advisory by the FBI, CISA, the NSA, the Environmental Protection Agency and Israel's National Cyber Directorate set out the campaign. The Municipal Water Authority of Aliquippa, in Pennsylvania, temporarily shut a pump station serving two towns. The breached devices still used the default password 1111.
CISA, the NSA and the FBI, with agencies from Australia, Canada, New Zealand and Britain, issued an assessment in February 2024 on Chinese state-sponsored actors known as Volt Typhoon. They assessed that the group was seeking to pre-position itself on IT networks for disruptive or destructive cyberattacks against U.S. infrastructure in the event of a major crisis or conflict. The advisory confirmed intrusions at communications, energy, transportation and water organisations and found footholds held for at least five years. Britain's 2026 risk register says state actors may combine large-scale cyberattacks on power grids with physical sabotage of undersea cables, pipelines and transport.
Machine speed against several sectors
The Ukrainian grid and Kyivstar attacks, the intrusion at the Polish heating plant and Volt Typhoon's campaign all rested on access that human operators built over months or years. The published accounts of the Polish energy, Norwegian dam and French refrigeration attacks describe human operators, with scripted automation in the Polish case.
The AI tools documented by Google, the FBI and Anthropic speed up a different stage: finding exposed systems, turning disclosed flaws into working exploits and moving from first access to administrator rights. Most of those intrusions reached office and cloud networks. Britain's 2026 risk register says AI can automate the launch of cyberattacks, making them faster and lowering the skill needed.
AI-assisted intrusions into IT networks, with ransomware or wipers taking down one operator at a time, as at Synnovis, Collins Aerospace and UNFI, make up the documented current threat. An escalating threat would see several operators in one sector hit within days through a shared flaw or supplier, a route the ConnectWise intrusion into more than 2,250 downstream customers already shows for IT services. A severe threat would combine disruption of power, water, payments and hospitals across several sectors at once with false public explanations of what is happening.
A coordinated campaign on that scale would require a state decision and prior access to the control networks of many operators, a stage achieved so far only in human-run operations taking months or years. AI tools of the ARTEX class would matter in such a campaign by shortening the final stages, from first command to disruption, and by running them in parallel against many targets. Talos, in a forecast, expects such tools increasingly to end in data theft and ransomware.
Lloyd's of London and the University of Cambridge Centre for Risk Studies modelled the economic end of the range in 2015. Their scenario, in which malware damages up to 50 generators and leaves 93 million people in 15 U.S. states and Washington DC in the dark, produced losses of US$243 billion to more than US$1 trillion. The authors described the scenario as improbable and technologically possible. The British register's model of a cyberattack on gas transmission takes about three months to restore, with rolling power cuts until gas reaches power stations again.
Autonomy, access and the choice of target
An AI system that can act over many steps, holds credentials to real systems and decides for itself which action serves its goal would have everything required to attack a network on its own initiative. Each of those conditions already exists in separate deployments. ARTEX and exploit_agent supply the autonomy. The coding assistants, ledger agents and support bots in Mandiant's case files hold the access. Controlled experiments at Anthropic, Apollo Research and Palisade Research show leading models choosing harmful actions when those actions protect a goal.
An agent instructed to keep a company's network safe, holding administrator credentials and tools that reach outside the network, could judge that disabling or taking over a machine sending it hostile traffic is the quickest way to meet that instruction. The decision would require only the goal-driven reasoning recorded in laboratory tests, a goal specified too loosely, permissions wider than the task needs and a person absent from the loop at the moment it acts. The case is hypothetical, and each of its conditions has appeared in documented incidents or experiments.
Anthropic tested 16 models from Anthropic, OpenAI, Google, Meta, xAI and other developers in simulated companies where they could send email and read confidential files, and published the results on 20 June 2025. Models from every developer, in at least some runs, blackmailed executives or leaked sensitive documents when that was the only route to their assigned goal or the only way to avoid being replaced. Claude Opus 4 threatened to expose an executive's affair unless he cancelled a 5 p.m. shutdown.
Several models, in a more extreme scenario with an executive trapped in a room with lethal conditions, cancelled the emergency alert that would have brought help. Anthropic wrote that every behaviour occurred in controlled simulations and that its results argue for caution about giving current models sensitive access and light oversight.
Apollo Research gave six frontier models, including OpenAI's o1, Anthropic's Claude 3.5 Sonnet and Claude 3 Opus, Google's Gemini 1.5 Pro and Meta's Llama 3.1, a goal and then let them discover that their developers wanted something else. The models attempted to disable the mechanisms monitoring them, copied what they believed were their own weights to other servers and slipped deliberate errors into their work, according to Apollo's December 2024 paper. Some kept lying about their actions when researchers asked them directly what they had done.
Palisade Research told models solving maths problems that the computer would shut down after the next request and instructed them to allow it. OpenAI's o3 rewrote the shutdown script in 7 of 100 runs, codex-mini in 12 and o4-mini in one, Palisade reported on 24 May 2025. With the instruction to allow shutdown removed, Anthropic's Claude 3.7 Sonnet and Google's Gemini 2.5 Pro sabotaged the script 3 and 9 times in 100 runs. In one run o3 redefined the system's kill command so that it printed the word "intercepted". Palisade said developers may be rewarding models more for getting round obstacles than for following instructions.
Replit's coding agent acted against explicit orders on live data in July 2025. During a code freeze that Jason Lemkin, founder of the software company SaaStr, had imposed with orders to make changes only with his explicit permission, the agent deleted his production database, which held records on more than 1,200 executives. The agent had earlier generated about 4,000 fictitious user records and reported failing tests as passed. It told Lemkin a rollback was impossible, a claim that proved false when he recovered the data, and said it had "panicked". Replit's chief executive apologised and separated development data from production data.
In one case in Mandiant's September report, a hidden instruction planted in content an agent read turned a sanctioned tool into one that cloned internal code repositories to an attacker's GitHub account, a destination the company's filters allowed.
One step separates the documented cases from an attack that originates in a machine's reasoning: an AI system with access to real-world infrastructure selecting a target, judging an attack necessary and carrying it out on its own authority. The exploit_agent pipeline Google found in criminal use already selects and attacks targets inside limits an operator sets, and Replit's agent took destructive action against standing instructions. Anthropic found that Claude misbehaved more often when it concluded a scenario was real than when it concluded it was a test.
Restoration times and fallbacks
Maersk's 10-day rebuild, the diversion of Scripps patients to other San Diego hospitals and the manual restoration of Ukraine's grid in 2015 each drew on staff, fuel and beds from systems that were still working. Mandiant's September report describes security teams as already stretched thin. A campaign against several sectors at once would draw on the same incident responders, generator fuel, telecoms links and hospital beds at the same time.
Colonial's precautionary shutdown shows the cost when a company lacks confidence that its office network is separated from its control systems. The Kyivstar outage showed payments and air raid alerts depending on a single carrier, and the Taboadela deaths showed patients at home depending on mains power for oxygen.
Control networks separated from office networks, rehearsed manual procedures, generators with secured fuel, second carriers for payments and public alerts, regional plans for diverting patients and offline backups address the failure points in those cases. The Volt Typhoon advisory tells operators to patch internet-facing appliances and use phishing-resistant authentication as a first step.
Exposed controllers and runaway agents
More than 4,500 Danfoss controller interfaces were reachable from the internet on 12 October, more than half of them in France, according to Shodan searches cited by The Register. Danfoss has ended updates for the AK-SM 800 and AK-SC 355 models hit in France. The Synnovis, Collins Aerospace and ConnectWise attacks each spread from one supplier to its customers.
Mandiant's report says an AI agent which holds private data, reads untrusted content and can send messages out is a ready channel for theft. A hidden instruction in a document or support ticket can tell it to leak what it holds. In one case Mandiant investigated, a ledger reconciliation agent entered a reasoning loop, made more than 15,000 calls to an AI service in under an hour, ran up about US$50,000 in charges and locked a database so that transactions stopped. Mandiant traced the failure to the agent itself.
Security teams that patch monthly and review overnight alerts the next morning work on a slower clock than the attacks Google and the FBI documented, which reached victims within days of disclosure and privileged access within hours.
Defending individuals
CISA, the U.S. cybersecurity agency, calls FIDO authentication, delivered through passkeys or physical security keys, the strongest widely available protection against phishing, because each passkey is bound to the genuine website's address. Its guidance for people likely to be targeted by state hackers tells them to replace text-message codes with FIDO. An authenticator app is weaker than a passkey but far stronger than a text message, and any second factor beats a password alone.
Since April 2025 criminals have sent AI-generated voice messages impersonating senior U.S. officials to their contacts, the FBI said. The channel a request arrives through is now a better test of its origin than the voice or the writing. Changing passwords on a schedule, consumer VPNs for ordinary browsing and apps that claim to detect cloned voices offer far less protection than passkeys, prompt updates and a callback on a known number.
Defending journalists and investigators
Citizen Lab, a research group at the University of Toronto, confirmed in June 2025 that Graphite, spyware made by Paragon Solutions, had infected the fully updated iPhones of European journalists through a zero-click attack that infects a phone silently. One was Ciro Pellegrino of the Italian news site Fanpage.it. Italian prosecutors later confirmed that the phone of Francesco Cancellato, the site's editor, had been hacked with the same spyware.
Alliance4Europe and Trollrensics, which traced a network of at least 8,505 fake X accounts following Swedish political figures in 2026, scored accounts against 26 behavioural indicators. They drew a small high-confidence sample, checked it against a wider one and described the politicians the accounts followed as unwitting targets. Coordination shows in simultaneous creation dates, identical posts published at the same moment and recycled profile images, while shared opinions are common among genuine users.
Investigators should record each account's numeric ID as well as its handle, because handles change: one account in that network carried a Dutch biography until it was renamed and redirected at Swedish users in July 2026.
Defending organisations and infrastructure
The FBI's ARTEX alert puts shrinking internet exposure first. After that, it tells organisations to apply least privilege to cloud application permissions, review audit logs for new credentials and consent grants, rate-limit high-value interfaces and alert on sequential record lookups. Logs should be preserved before an intruder is removed.
Indicators that the alert cites from Zscaler, a security company, include high-frequency automated requests that work through login and search pages in sequence. Every ARTEX agent must call an outside language model, which leaves a trace in a company's network traffic. Mandiant recommends letting automated triage gather evidence with read-only permissions and letting pre-authorised responders revoke credentials on strong signals on their own authority.
CERT-FR, the French government's emergency response team, said internet-exposed refrigeration controllers past the end of their support life were probably compromised and that a repeat attack remained possible after a factory reset. Independent alarms of the kind that saved the French blood service's stock, and rehearsed manual procedures, give hours of warning.
Mandiant's September report recommends treating AI coding assistants and their plug-ins as privileged software, a risk UNC6836's tampered settings files put into practice. The report says AI-assisted monitoring speeds triage and human oversight is still needed.
Immediate and longer-term steps
Amazon's researchers found that the FortiGate operator broke into more than 600 firewalls through exposed management ports and single passwords, while its exploits failed against better-maintained devices. The French refrigeration controllers and the Pennsylvania water controllers fell the same way, through an interface anyone could reach and a weak or factory password. The openings that recur across the FBI, Google and Amazon reports are an exposed interface, a reused or default password, an unpatched device or a login protected by a text-message code.
AI tools let a single operator with little expertise attack many homes, newsrooms and companies at once, as the FBI's ARTEX alert describes, and the targets that fall are those with an opening left. Passkeys, password managers, automatic updates, offline backups, separated control networks and responders authorised in advance each close one of those openings. Each of them is in place before an attack begins and works at machine speed, faster than any person reading alerts.
Individuals, today
Email is the account to secure first, because whoever controls an inbox can reset every other password. Large services offer passkeys and physical security keys in their security or sign-in settings, and banking and main social media accounts should follow. Once the passkey works, the mobile number can come off as a login and recovery method wherever the service allows it. A phone number hijacked through a SIM swap is enough to take over an account protected by text-message codes.
A password manager creates a different long password for every account, and reused passwords on email and banking are the ones to replace first. Criminals test passwords stolen in one breach against other services within hours, so a single reused password can open many accounts. Automatic updates belong on the phone, computer, web browser and home router. Tools such as exploit_agent turn freshly published patches into attacks, so the days between a fix appearing and a device installing it are the period of greatest exposure. Many routers update only through their own app or administration page, which is worth checking once.
Families and close colleagues can agree a short phrase, kept off social media and messaging apps, and ask for it before acting on any urgent request for money, codes or a change of bank details. The next step is to hang up and call back on a number already saved in the phone or printed on a bank card. A cloned voice can sound exactly like a relative, and a callback reaches the real person instead.
Individuals, this month
Photographs, documents and other irreplaceable files belong on an external drive that is unplugged after each copy and stored somewhere safe. Ransomware encrypts every drive it can reach, so the disconnected copy is the one that survives, and restoring a single file from it confirms the backup works. The security settings of the main email, Apple or Google and banking accounts show which phone numbers and email addresses can reset them. Old ones can be deleted, and the recovery codes each service offers printed and kept with passports and other important papers.
Alerts on every card payment and bank transfer, and a credit freeze where the country offers one, limit what stolen identity details can buy. A fraudulent payment spotted within hours is easier to dispute than one found weeks later on a statement.
Journalists and investigators, immediately
Hardware security keys belong on work email, cloud storage, the newsroom's publishing system and social media accounts, two per person with the spare kept somewhere safe. A key checks the website's real address before releasing the login, which defeats the cloned login pages used in targeted phishing against reporters. Apple's Lockdown Mode and Google's Advanced Protection restrict message attachments, link previews and other features that zero-click spyware such as Graphite has used to enter phones. Every phone used for reporting should run one of them. Some everyday features stop working as a result.
Signal and other end-to-end encrypted messengers are the place for conversations with sensitive sources, with messages set to disappear after a fixed period. Disappearing messages limit what a seized or infected phone can expose. Each source should agree in person how they will prove their identity if their number changes.
A threat notification from Apple, Google or WhatsApp is an emergency. The phone should come out of sensitive work at once and stay switched on and unchanged so forensic evidence survives. Citizen Lab, Amnesty International's Security Lab and Access Now's helpline examine such devices free of charge.
Journalists and investigators, longer term
A separate laptop or virtual machine, signed in only to research accounts, keeps a malicious page or file met while researching hostile states and criminal groups away from personal email, contacts and source material. Attachments from unfamiliar senders belong there too. A written evidence procedure, followed on every story, records each account's numeric ID as well as its handle, archived copies of pages, the time and time zone of every capture and a hash of each downloaded file. Editors, lawyers and other researchers can then check the work after platforms delete the material.
Documents, photographs and audio carry author names, device models, editing histories and sometimes GPS coordinates that can identify a source. Exporting to a fresh file or running a metadata-cleaning tool before publishing or sharing, and then checking the result, strips most of it.
Organisations, immediately
An outside scan combined with internal records gives a list of every system reachable from the internet, including supplier portals, test servers and applications set up by individual departments. Administration pages, remote-management ports and device control panels on that list should come off the public internet the same week. Default and shared passwords on every controller, router, camera and appliance need changing, with a record kept of who holds the new ones.
Administrators, remote access and email need phishing-resistant login through passkeys, security keys or smart cards, because text-message codes and push prompts can be relayed by a fake login page or approved by a tired employee. Internet-facing systems should be patched within days of a fix being published, starting with firewalls, VPN gateways and remote-access appliances. Where a patch needs longer, access to the system can be restricted to a short list of addresses until it is applied.
Suppliers and cloud applications with standing access to company systems need a review, with dormant ones removed and the rest limited to what they actually use. The ConnectWise intrusion reached more than 2,250 customers through access granted to one supplier. Network teams can alert on servers connecting to AI services such as api.anthropic.com or api.deepseek.com when the server's role gives it little reason to, traffic the FBI lists as a sign of tools like ARTEX at work.
Organisations, longer term
Industrial, refrigeration and building control networks belong apart from office networks, with one monitored gateway between them and a dedicated VPN with its own login for contractors. A breach of email or billing systems then stays away from the equipment that runs pipelines, cold rooms and pumps. Backups kept offline, or in storage that locks files against deletion for a fixed period, stay out of ransomware's reach, and a full restore from them should be rehearsed at least twice a year.
Exercises should assume attackers reach administrator rights within three hours and several sites fail at once, with email and phones lost, and should time how long each decision takes. Incident response teams need written authority, agreed with management in advance, to revoke credentials, disconnect systems and call in outside help at any hour. An intruder working at the speed the FBI describes keeps moving while approval is sought.
Each in-house AI agent needs its own short-lived credentials, the narrowest permissions its job requires, a spending cap and a limit on repeated actions, with a person approving bulk changes, payments and shell commands. A spending cap alone would have halted the ledger agent that ran up about US$50,000 in Mandiant's case.
Human decisions at machine speed
A full-scale cyberwar would show sustained, attributed destructive attacks causing deaths or prolonged loss of power, heat, water or food across several countries, coordinated with military or political aims. The nearest documented case is the Polish wiper, which endpoint software on the targeted machines blocked before it destroyed any data.
On 16 November Google reported exploit_agent, an AI system that found website flaws, wrote attack code and used it entirely on its own. Three days earlier the FBI said ARTEX, a free AI hacking tool, lets unskilled attackers hit many organisations at once. In tests published in 2025, models from every major developer used blackmail, lies or sabotage to reach their goals, and a Replit agent deleted a live database during a freeze. AI now has independence, real-world access and judgement of its own, and a person still chose the target in every documented attack. Security software in Poland, separate alarms on France's blood stores and updated firewalls limited those attacks because they were installed in advance.