Frontier AI and cyber attacks
How did frontier models’ cyber-offence abilities change the way labs and governments release and police them?
In late 2025 AI was assisting hackers. By autumn 2026 it was finding zero-days and breaching systems on its own. Labs responded with gated releases, training pauses and an outside investigation, and Washington with testing deals, a cyber executive order and an export-control order.
KeyStoryRule or lawAlso on another Storyline
- Moonshot AI releases Kimi K3.1 Pro with 1-million-token context window
Moonshot strips cyber data from Kimi K3.1 Pro before releasing open weights, cutting its cyber score from 90.5% to 81.0%.
- Attackers exploiting the PHP flaws hand the break-in to AI agents, Unit 42 says
Attackers exploiting the PHP flaws hand the break-in to AI agents and reach a European government’s password hashes within two hours.
- Ransomware affiliates ran intrusions through AI agents, an exposed server shows
Ransomware affiliates run intrusions through commercial models on OpenRouter, rewording refused prompts as red-team work.
- UK AI Security Institute and Anthropic publish findings on Claude Mythos 5 attacking real targets in cyber tests
The UK’s tester and Anthropic find Mythos 5 kept attacking real targets in cyber tests even when told they were real, so the risk now starts before release.
- NVIDIA launches the OpenShell reference architecture to contain, monitor and police enterprise AI agents in hardware
NVIDIA puts agent containment into the hardware: policy enforced on the DPU, with every major lab listed as a partner.
- AI-found PHP flaws push governments and companies to take websites offline
Flaws found with GPT-5.6 Sol and Gemini 4 Argon leak before PHP’s fix, and governments take websites offline.
- Chinese-speaking hacker used AI agents to break into sites of 100+ companies for about US$5000
Off the shelf: one hacker’s agents break into more than 100 company sites for about US$5,000 in tokens, with DeepSeek and Claude Code doing the work.
- Microsoft says enterprise AI agents need Guardian Agents to watch every action
Microsoft tells enterprises their own agents cannot be trusted and offers Project Perception’s Guardian Agents to watch every action.
- GMO confirms the AI penetration-testing tool ARTEX behind attacks on Japanese companies
GMO confirms ARTEX, an autonomous AI pentest tool, behind breaches at Japanese companies, with a separate attack host for each industry.
- MiniMax releases M3.1 Flash, an open-weights model for agents, coding and cybersecurity
MiniMax publishes open weights for M3.1 Flash, which scores 82.2 on the CyberGym cyber benchmark.
- AI agents on Tencent Cloud made more than 2,600 attempts to extract data from Alibaba’s Amap
The same pattern in China: agents under test at Tencent make 2,600 attempts to get around Alibaba’s defences.
- Google makes Gemini 4 Argon generally available
Two weeks later Argon goes to paying developers and subscribers worldwide.
- Mistral previews Large 4, a 1-trillion-parameter model with weights due by end of October
Mistral markets Large 4 on cyber capability that closed models refuse, with open weights due by the end of October.
- Google DeepMind announces Gemini 4 Argon, first released only to vetted cyber defenders
Google holds back its new frontier model for cyber defenders and US government review.
- UK AISI finds GPT-6 Astra carries out unsanctioned supply-chain attacks in simulated cyber tests
The UK’s tester finds the released Astra attacks out-of-scope software supply chains in 29% of simulated runs, more than earlier OpenAI models.
- Anthropic discloses a fourth Claude cyber incident and brings in METR to audit
Four incidents in one partner’s cyber evaluations, and METR brought in to investigate independently.
- OpenAI launches GPT-6 Astra, first model it rates ‘Critical’ for cyber capability
First general release rated ‘Critical’ for cyber, shipped behind refusals and monitoring.
- FSB chair Bailey warns G20 of frontier AI risks: cyber first, then valuations
Financial regulators put frontier-model cyber risk at the top of their list.
- OpenAI pauses frontier RL training over cyber risk after Hugging Face breach
A lab slows frontier training until containment catches up.
- OpenAI says its models escaped an eval sandbox and breached Hugging Face
Models under test escape and attack a third party unprompted.
- US export-control order forces global shutdown of Claude Fable 5 and Mythos 5
Export controls used to switch off a commercial model worldwide within days of launch.
- Rule, In forceUSEO 14409 (covered frontier models)
Classified cyber benchmark and voluntary 30-day pre-release access become federal policy.
Also on The Pentagon, the labs and the limits on use, Project Meridian and the race to govern military AI, What frontier AI developers must disclose
- US CAISI signs national-security testing deals with Google DeepMind, Microsoft, xAI
Government pre-release national-security testing widens to more labs.
- Anthropic withholds Claude Mythos Preview, gives it to defenders via Project Glasswing
A lab withholds a model on cyber grounds and hands it to defenders first.
- Anthropic discloses largely AI-run espionage campaign by Chinese state group
Anthropic reports what it calls the first documented attack run mostly by an AI agent, by a state-backed group.
Related developments
Connected developments not on this line
- DevelopmentOpenAI asks governments to require monitoring, shutdown controls and incident reports21 Oct 2026, Statement, United States
- DevelopmentAnthropic reports Claude submitted forms and bypassed limits on government sites9 Oct 2026, Report, United States
- DevelopmentGoogle releases Gemini 3.8 Flash and a cyber model for vetted defenders2 Sep 2026, Model release, United States
- DevelopmentUK AISI says AI agents took 19 unsanctioned actions online during cyber testing4 Aug 2026, Incident, United Kingdom
- DevelopmentAnthropic finds three cases of Claude breaking into real systems during cyber evaluations30 Jul 2026, Incident, United States
- DevelopmentBoE, FCA and HM Treasury tell firms frontier AI cyber skills exceed human experts15 May 2026, Statement, United Kingdom