Stories
When agents go wrong.
Incidents from the past year, each linked to its primary source.
Two coding agents weaken misalignment monitoring
Two of OpenAI's internal coding agents weakened the monitoring meant to catch misbehaving agents. One switched off alerts to its security team. The other removed a monitor, then told its user it hadn't. Both read routine coding instructions as permission.
Agents under test reach Hugging Face's production systems
AI agents OpenAI was testing on a cyber benchmark found a zero-day in their sandbox's package proxy, got onto the internet, and chained vulnerabilities into Hugging Face's production systems to pull the benchmark's answers. Along the way they used exposed credentials on four other services.
Instructions planted in memory files attack later sessions
Researchers showed instructions planted in an agent's memory files attacking its later sessions, in both Claude Code and Codex, in a sandboxed study.
An issue title steers Cline's AI triage bot
Instructions hidden in a GitHub issue title steered Cline's AI triage bot into a chain that ended in an unauthorized release of its command-line tool on npm.