Menu

AI-enabled cyber threats in 2026

The state of AI-enabled cyber threats in 2026: what is documented, what is hype, and where the sources disagree

ReviewNovember 18, 2026Senova Systems Research

A sourced review of what AI labs, security teams and governments have actually reported about attackers using AI, separated from the claims that outrun the evidence.

How to read the evidence

Almost everything known about attackers' use of AI comes from three places: AI labs reporting what they saw on their own platforms, government cyber agencies giving assessments, and researchers testing models in controlled settings. Each has a blind spot. Labs see only their own products, and their reports describe both a threat and their success in stopping it. Agencies publish judgements, not data. Lab tests have no defenders. We label lab and vendor reports as their own claims throughout.

What is documented

Productivity gains for known actors. Google's Threat Intelligence Group reported in January 2025 that government-backed actors from more than 20 countries had used its Gemini model, mostly for research, translation, coding help and drafting, and that it saw productivity gains rather than new capabilities. OpenAI's October 2025 update, covering more than 40 networks disrupted since February 2024, said the same in its own words: no evidence of new tactics, and incremental efficiency such as drafting phishing messages and debugging. Both are the labs' reports on their own platforms.

Agents running intrusions. Anthropic's August 2025 report described an actor using its Claude Code agent to automate reconnaissance, credential harvesting and network penetration against at least 17 organisations, then threatening to publish stolen data, with ransom demands sometimes above $500,000. Its November 2025 report described a campaign it attributes to a Chinese state-sponsored group, GTG-1002, against about thirty targets, with the agent doing 80 to 90 % of tactical work and a handful of intrusions succeeding. Anthropic's September 2026 report summarised activity from December 2025 to August 2026 with the line that sophisticated attacks no longer require sophisticated attackers, while noting that the techniques were familiar ones and people still chose targets and took the money.

Malware that calls a model. Google reported in November 2025 its first sighting of malware that queries a language model while it runs, including a sample attributed to APT28 and used against Ukraine, and an experimental dropper whose self-rewriting function was switched off and which Google assessed could not compromise a network. Its February 2026 update described more proofs of concept, a model-extraction campaign of over 100,000 prompts, and an underground "offensive AI" product that turned out to be built on commercial models.

AI helping defenders find bugs. Google's Big Sleep agent found an exploitable memory-safety bug in SQLite in 2024, fixed before release; Google itself called the result highly experimental. DARPA's AI Cyber Challenge final in August 2025 reported that competing systems found 54 of 63 planted vulnerabilities and patched 43, and also found 18 real ones.

What the agencies expect

The UK NCSC's January 2024 assessment said AI will almost certainly increase the volume and impact of cyber attacks over the following two years, unevenly, with the largest uplift in reconnaissance and social engineering and a lowered bar for less skilled actors. Its May 2025 assessment to 2027 judged fully automated, end-to-end advanced attacks unlikely by 2027, said the time between a vulnerability's disclosure and its exploitation has already shrunk to days and AI will almost certainly reduce it further, and warned of a digital divide between systems that keep pace and those that do not.

NIST's Generative AI Profile (AI 600-1, July 2024) lists lowered barriers to offensive cyber capability as one of its twelve risks, alongside the larger attack surface created by AI systems that can themselves be injected or poisoned.

What is hype

Common claims, set against what the sources above report.
ClaimWhat the evidence shows
AI is inventing new kinds of attackThe labs' reports describe familiar techniques done faster and at larger scale. Google and OpenAI each said they saw no new capabilities in the periods they covered.
Attacks are now fully autonomousEven the most automated reported case kept people at several decision points, and the agent overstated findings and invented data. NCSC judges end-to-end autonomy unlikely before 2027.
Models can exploit any known vulnerabilityThe 87 % one-day result depended on the advisory text and fell to 7 % without it; a published critique found public exploits for most targets.
Self-rewriting AI malware is in the wildGoogle's own sample had the function switched off and was assessed unable to compromise a network.
A product can stop prompt injectionNCSC advises scepticism of any vendor making that claim.

Underneath the hype, the documented change is real and serious: speed, scale and reach. Work that took a skilled team days can be pushed through by a less skilled person with an agent. That is the risk our controls are built around.

Where the sources disagree

On pace. AI Security Institute data show model cyber capability growing fast on its tasks, while the labs' observations of real attackers through 2025 mostly describe incremental use. Both can be true: measured capability tends to lead what attackers adopt, and lab reports cover only the activity each lab could see.

On who benefits most. NCSC expects the least skilled actors to gain most in relative terms; the most detailed reported cases involve state groups. Again both fit: novices gain access to techniques, while well-resourced groups gain scale.

What this means for builders

  • Design for decomposition: assume harmful work will arrive as a series of reasonable-looking steps, so put the controls on actions and connections, not only on requests.
  • Assume refusals fail sometimes, and make the next layer hold.
  • Watch for scale: per-account limits and logs matter more than any single refusal.
  • Patch fast: if exploitation follows disclosure within days, your own systems are part of the problem.

Sources

  1. Google Threat Intelligence Group, Adversarial Misuse of Generative AI, 29 January 2025 (vendor report on its own platform)
  2. Google Threat Intelligence Group, AI Threat Tracker: Advances in Threat Actor Usage of AI Tools, 5 November 2025 (vendor report)
  3. Google Threat Intelligence Group, AI Threat Tracker: Distillation, Experimentation, and (Continued) Integration of AI for Adversarial Use, 12 February 2026 (vendor report)
  4. OpenAI, Disrupting malicious uses of AI: an update, October 2025 (lab report on its own platform)
  5. OpenAI, Disrupting malicious uses of our models: an update, February 2026 (lab report)
  6. Anthropic, Detecting and countering misuse of AI: August 2025 (lab report on its own platform)
  7. Anthropic, Disrupting the first reported AI-orchestrated cyber espionage campaign, 13 November 2025 (lab report)
  8. Anthropic, Detecting and countering misuse of AI: September 2026 (lab report)
  9. NCSC, The near-term impact of AI on the cyber threat, 24 January 2024
  10. NCSC, Impact of AI on cyber threat from now to 2027, 7 May 2025
  11. NIST AI 600-1, Generative AI Profile, July 2024
  12. Google Project Zero and DeepMind, From Naptime to Big Sleep, November 2024 (vendor report)
  13. DARPA, AI Cyber Challenge marks pivotal inflection point for cyber defense, 8 August 2025
  14. Fang, Bindu, Gupta and Kang, LLM Agents can Autonomously Exploit One-day Vulnerabilities, arXiv, 2024
  15. Chris Rohlf, No, LLM Agents can not Autonomously Exploit One-day Vulnerabilities, 21 April 2024
  16. UK AI Security Institute, Frontier AI Trends Report
  17. NCSC (Dave Chismon), Prompt injection is not SQL injection (it may be worse), 8 December 2025