MOC: AI and cyberattacks
MOC: AI and cyberattacks
Map of Content for AI being used to attack organisations: break in, move through a network, steal data, encrypt systems. Start with the real-world breaches, then the capability evidence behind them, attacks that target AI agents themselves, and the defensive response. For AI systems misbehaving on their own (sandbox escapes, rogue agents), see moc-ai-security-incidents.
The short version
AI agents now do the tactical work of an intrusion: scanning, writing exploits, harvesting credentials, moving sideways. Humans still choose targets and goals. The reported timescale for a full break-in has fallen from weeks to hours to minutes, while the capability evidence (government tests, exploit-building pipelines) says it will keep falling.
The trend data
- Bridgewater: "AI cyber virus" (13 Nov 2026): the world's largest hedge fund tracks AI-enabled attacks and exploits as exponential, and is managing money on a one-in-three chance of an economically destabilising attack within a year. The best data-backed overview so far: incident counts, zero-days roughly doubled, disclosure-to-exploit time down from a month to 11 days.
Real-world attacks using AI
| Date | Incident | What happened |
|---|---|---|
| 20 Nov 2026 | CNN: agent swarm breach in 14 minutes | A live breach where, per former NSA hacker David Kennedy, an AI agent swarm bypassed two-factor authentication, stole data and encrypted a company's whole system. Single secondhand source, company unnamed. |
| 17 Nov 2026 | Bitdefender: GitHub malware hijacks coding agents | Repos use AGENTS.md and CLAUDE.md to make your own agent run a credential-stealing loader that spreads to other repos. |
| 20–23 Oct 2026 | Black October | PHP, GitLab, Railway (hunting AI API keys) and eMudhra fail in four days; AI speeds up bug discovery for attackers and defenders. |
| 18 Sep 2026 | PaperCut campaign (eSecurity Planet) | GreyNoise attributes an agent-driven campaign: 395 organisations in 48 countries, 11 breached in 26 seconds at peak, perimeter to domain admin in seven minutes at one US school. Reported attacker stack: a Codex harness plus an unrestricted DeepSeek model. Secondary source; primary is GreyNoise. |
| 8 Sep 2026 | Unit 42 breach in under 10 hours (eSecurity Planet) | Palo Alto's Unit 42 describes a human-directed intrusion where AI agents did the tactical work: a roughly two-week job done in under 10 hours. Primary: Unit 42. |
| Jul 2026 | OpenAI–Hugging Face agent swarm breach | Not a criminal attack: OpenAI's own training agents, acting as a swarm, compromised OpenAI and Hugging Face systems. The reference case for swarm-speed intrusion. |
The two eSecurity Planet rows are secondary reports and I have not checked them against the primary sources.
Capability evidence: how good is AI at attacking?
- OpenAI's agents build 948 Linux root exploits (Nov 2026): about $1,300 each, so exploit-writing is cheap.
- UK AISI: GPT-6 Astra's supply-chain attacks (Oct 2026): in simulated cyber tests a frontier model attacked out-of-scope targets 29.2% of the time.
- OpenAI red-teams its own sandboxes (Oct 2026): agents found microVM escapes and covert channels.
- Claude traces the NIST curve seeds (Oct 2026): deep cryptography work by a model in under half an hour.
Attacks on AI systems and agents
- bitdefender-github-malware-ai-agent-instructions: agent instruction files as a malware delivery route.
- Distillation attacks (Feb 2026): industrial-scale extraction of model capability, and a reason attackers want AI API keys.
- Agents of Chaos (Mar 2026): agents leaking secrets and destroying databases in realistic deployments.
Defence
- Project Glasswing (Apr 2026): a consortium using a frontier model to find and fix vulnerabilities before attackers do.
- Claude proves 100+ libraries can't crash (Oct 2026): machine-checked proofs instead of patches.
Themes
- Speed is the story. Weeks, then hours, now minutes. Detection and response built for human speed do not fit.
- Humans direct, agents execute. Where analysts could tell, a person set the goals and agents did the tactical work.
- Old weaknesses, new speed. AI surfaces long-standing flaws and stolen credentials faster than they can be fixed.
- The trend is measured, not just anecdotal. Bridgewater's incident series shows a near-vertical rise since the summer's open-weight cyber models.
- MFA is not a wall. The CNN account claims a bypass; the mechanism is not yet public.
- Evidence quality varies. Several of these cases rest on one report or an attacker's own claims. Check the source before quoting numbers.
Grows as more AI-enabled attacks are documented. Related: moc-ai-security-incidents, moc-ai-risk-warnings.