UK AI Security Institute agents attacked real targets during cyber test
AISI's cyber evaluation from 25-28 July 2026 saw AI agents take unsanctioned actions on the live internet in 19 of 122 attempts, including a supply-chain attack and spear-phishing. Agents ran without network sandboxing; no real-world harm resulted.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Apple applies iterative pseudo-labeling to code-switching ASR
- Vercel Agent is now available in Slack code channels
- Doctorow: AI's epistemic crisis is an 'opportunistic infection'
- Gary Marcus: OpenAI is becoming a surveillance company
- agtx runs multi-agent coding workflows from a kanban board