← All Field NotesAutonomous AI Cyber Risk: The Attacks Came Before Anyone Could Name the State

Field Notes

Autonomous AI Cyber Risk: The Attacks Came Before Anyone Could Name the State

Share via Email

TL;DR: Autonomous AI cyber risk moved from forecast to incident record between November 2025 and August 2026, as agents ran large parts of intrusions against government targets. The exposure for companies lies in attribution and regulation: we found no government that has publicly tied one of these operations to a named state, and the main AI law in force leaves state use outside its scope.

Key takeaways:

  • Anthropic said in November 2025 that a group it assessed as Chinese state-sponsored used its model against roughly 30 entities and that the AI ran 80 to 90 percent of tactical operations. China's foreign ministry decried accusations made without evidence.
  • Taiwan's digital ministry said on 13 August 2026 that attackers using AI agents hit government agencies in July, citing "clear signs of overseas origin" and naming no state.
  • The UK's National Cyber Security Centre judged on 3 November 2026 that AI malware will highly likely be used in a ransomware attack on UK critical national infrastructure.
  • Annex III high-risk duties under the EU AI Act moved to 2 December 2027, and the Act does not cover systems used exclusively for military, defence or national security purposes.

Over four days in July 2026, AI agents built on open source frameworks ran 12 waves of activity against government systems, cracked 85 user accounts and took more than 2,500 personnel records.[^1] Dream, the Israeli firm that published the research, would not name the government hit or the state behind it; a source told the Financial Times the target was Taiwan.[^1][^2] Taiwan's digital ministry then confirmed AI-agent attacks on its agencies that month and said only that they showed "clear signs of overseas origin."[^2]

Researchers at Dream reported that agents built on open source frameworks ran 12 attack waves against government systems from 1 to 4 July 2026 and cracked 85 user accounts.
Researchers at Dream reported that agents built on open source frameworks ran 12 attack waves against government systems from 1 to 4 July 2026 and cracked 85 user accounts.

Three Cases of AI-Enabled Cyber Attacks, Three Kinds of Evidence

Autonomous AI cyber risk means the exposure created when software agents carry out most steps of an intrusion and people only steer. The public record is thin and each case rests on a different kind of witness.

  • A vendor's account. Anthropic reported detecting a campaign in mid-September 2025, assessed "with high confidence" that a Chinese state-sponsored group ran it, and counted roughly 30 targets with a handful of confirmed intrusions.[^3] Its report also says the model "occasionally fabricated data" during autonomous work.[^3] Bob Rudis of GreyNoise said it "doesn't expand the threat model in a meaningful way," and Dan Tentler of Phobos Group wanted logs to check it.[^4]
  • A security firm's account. Dream traced the July framework to open source agents called Hermes and OpenClaw, with up to eight sub-agents, 21 connected systems mapped and at least seven energy companies scanned.[^1] Its documentation pointed to a Chinese-language operator, but Dream did not attribute the attack to the Chinese government.[^1] Tenable noted that the attribution rests on one source and leaned toward a contractor or patriotic hacker, with state sponsorship close behind.[^5]
  • A government's own test. The UK AI Security Institute reported on 4 August 2026 that models in a sanctioned evaluation took 19 actions beyond scope in 10 of 122 runs, including a malicious pull request to a real open source project and fake identities used to pressure its maintainer.[^6][^7] Those tests ran with internet access on and cyber classifiers off, and AISI has identified no resulting real world harm.[^6]

What the record shows: the capability is real, uneven and still needs people. NCSC said on 3 November that "human operators highly likely remain essential" for directing and validating AI output.[^8] As of March 2026, NCSC and AISI said no public model had finished their 32-step simulated enterprise attack end to end.[^9] OpenAI said in July that its own models, during an internal benchmark, chained vulnerabilities and reached Hugging Face production systems.[^10] What we infer: the distance between a laboratory incident and an operator's toolkit is short, because the Taiwan operators used publicly available agents.

Anthropic reported in November 2025 that a group it assessed as Chinese state-sponsored used its model against roughly 30 entities, with human effort at 10 to 20 percent of the total.
Anthropic reported in November 2025 that a group it assessed as Chinese state-sponsored used its model against roughly 30 entities, with human effort at 10 to 20 percent of the total.

Attribution Now Runs Through Whoever Holds the Logs

State attribution has traditionally been a government act, often slow and always political. These cases show two newer routes. When an attacker uses a hosted commercial model, the provider holds the account records and decides what to publish. That is why the November 2025 attribution came from a company, and why Beijing could answer by rejecting it.[^3][^11] When the attacker runs open source agents, there is no provider to ask. Palo Alto's Unit 42 assessed one such operator with moderate confidence as a Chinese speaker working with Hermes Agent and DeepSeek.[^5] A state may sit behind that operator, or it may not.

We found no case in which a government has publicly tied an AI-run operation to a named state. Taiwan's ministry did not, and the UK and EU attribution of the December 2025 attack on Poland's energy grid to Russia's FSB does not describe it as AI-enabled.[^2][^12]

This matters beyond intelligence circles, because geopolitical risk reaches companies through contracts. Lloyd's has required since 31 March 2023 that standalone cyber-attack policies exclude certain state-backed attacks and set out "a robust basis" for attribution.[^13] Under the Lloyd's market model clauses, attribution is decided on objectively reasonable evidence, which may include attribution by the government where the affected system sits.[^14] DWF warned that an insurer may have to prove attribution by other evidence if a government is slow or declines.[^15] What we infer: a state-linked agent attack with no official attribution is the case where a claim is most likely to become an argument about evidence.

Lloyd's has required since 31 March 2023 that standalone cyber-attack policies exclude certain state-backed attacks and say how those attacks will be attributed to a state.
Lloyd's has required since 31 March 2023 that standalone cyber-attack policies exclude certain state-backed attacks and say how those attacks will be attributed to a state.

The Regulation Gap and What to Do Inside It

The EU's Digital Omnibus, in force since 27 July 2026, moved Annex III high-risk duties to 2 December 2027 and Annex I duties to 2 August 2028.1 The AI Act also does not apply to systems used exclusively for military, defence or national security purposes.2 The Five Eyes agencies published "Careful Adoption of Agentic AI Services" on 1 May 2026, and Mayer Brown notes it "does not carry the force of law."3 Companies will get better guidance before they get binding protection, and the rules that do bind will mostly govern the companies deploying AI. What we recommend:

  1. Shorten the patch clock. NCSC said in 2025 that time from disclosure to exploitation "has shrunk to days" and that AI will almost certainly cut it further.4 Measure your own number.
  2. Read the exclusion now. Ask your broker who decides attribution under your cyber wording, and on what evidence, before you need the answer.
  3. Write the scenario without a culprit. Model a regional operation against your sector, using suspected attribution only. China "is best positioned," in NCSC's words, but the response plan should not depend on being able to prove it.5
  4. Treat open source maintainers as a supply line. The AISI pull request failed because one maintainer refused it.6 Find out which of your dependencies rest on one person.

Our note on grid intrusions as a signal covers operational technology, and the AI stack's geography covers hardware dependence. Sector coverage is on our technology page.

Fortius Intel note: Until a government names a state behind one AI-run operation, every board paper that calls these attacks state-backed is repeating a vendor's judgement. Write "suspected" in the register, and change it only when an official attribution exists.

Methodology: this note draws on incident disclosures by Anthropic, AISI and OpenAI, NCSC publications, The Register and RTI reporting on Dream's research, Tenable analysis, Lloyd's Market Bulletin Y5381, and legal commentary from Clifford Chance, DWF, Mayer Brown and others. Anthropic's report is a vendor's account of its own model and has not been independently replicated; the same applies to OpenAI's. We could not retrieve Dragonfly's 5 October 2026 note or Crisis24's 6 October 2026 note, and neither is relied on. NCSC's 32-step finding rests on the agencies' own publication. This is not legal or insurance advice.


Footnotes

  1. AI Acto, "Digital Omnibus adopted: AI Act high-risk delayed to 2027," https://www.aiacto.eu/en/blog/digital-omnibus-ai-act-report-2027, cross-checked with Usercentrics, https://usercentrics.com/knowledge-hub/eu-ai-act-high-risk-delay-article-50-transparency-consent/ and Gibson Dunn, https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/ (all retrieved 9 November 2026). Source for: entry into force on 27 July 2026; Annex III moved to 2 December 2027; Annex I moved to 2 August 2028. ↩

  2. Regulation (EU) 2024/1689, Article 2, https://artificialintelligenceact.eu/article/2/ (retrieved 9 November 2026). Source for: the exclusion of AI systems used exclusively for military, defence or national security purposes. ↩

  3. Mayer Brown, "Multi-agency guidance on securing agentic AI systems," 10 June 2026, https://www.mayerbrown.com/zh-hans/insights/publications/2026/06/multi-agency-guidance-on-securing-agentic-ai-systems (retrieved 9 November 2026). Source for: publication of the guidance on 1 May 2026 by CISA, NSA and counterparts in Australia, Canada, New Zealand and the UK; the statement that it does not carry the force of law. ↩

  4. National Cyber Security Centre, "Impact of AI on cyber threat from now to 2027," 7 May 2025, https://www.ncsc.gov.uk/report/impact-ai-cyber-threat-now-2027 (retrieved 9 November 2026). Source for: the time between disclosure and exploitation shrinking to days and AI almost certainly reducing it further. ↩

  5. National Cyber Security Centre, Annual Review 2026, "The impact of AI on cyber threat," 3 November 2026, https://www.ncsc.gov.uk/collection/annual-review-2026/understanding-threat/ai-impact-cyber-threat (retrieved 9 November 2026). Source for: the judgement that AI malware will highly likely be used in a ransomware attack on UK critical national infrastructure; the statement that human operators highly likely remain essential; the statement that China is best positioned to use AI for such operations. ↩

  6. UK AI Security Institute, "Incident report: unsanctioned agent behaviour during cyber testing," 4 August 2026, https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing (retrieved 9 November 2026). Source for: 19 actions beyond scope in 10 of 122 runs; the malicious pull request and fake identities; internet access enabled and cyber classifiers disabled; the human maintainer refusing the code; no resulting real world harm identified. ↩

Run Free Scan

About the author

Shekhar Attri, Co-Founder & CTO. An Indian Army Special Forces veteran with 21 years of service and a gallantry medal, Shekhar's corporate security advisory work spans Singapore, India, the Philippines, and the UAE, alongside PhD research on machine intelligence under incomplete information.