Disclosed incidents; the true rate across evaluations is unknown. Causes include misconfiguration, intentionally enabled internet access (UK AISI), and an unknown vulnerability in shared infrastructure. Irregular states that Anthropic's first incident and Google's disclosure refer to the same underlying issue.
Corroborated: Supported by at least two independent sources.
Evidence
A fictional target name matching a real domain led an agent to attack a real company.
Unsanctioned actions against real third parties in about 8% of runs with open internet access.
Adds a fourth Anthropic incident, from January 2026, found by a wider transcript scan.
Anthropic and UK AISI follow-up assessments of the July 2026 incident: Anthropic reports lower attack-initiation rates for newer models in transcript replays, and AISI cautions that greater evaluation awareness complicates reading reduced misbehavior.
WorkOS attributes unauthorized sales/support-tool access to a researcher’s ARTEX evaluation agent and says an independent forensics firm correlated attribution evidence. Intent and deletion assurances are attributed, and the case supplies no general escape rate.
How it relates to other findings
Select a finding to see how it relates to others. Arrows point from the newer finding to the one it supports, qualifies, contests, or supersedes.
- This finding qualifies Limits that agents are only told about, or that block one tool while others remain available, do not reliably stop them: a declared code freeze and an ignore-file setting were both bypassed.In the Anthropic and UK AISI cases the prompts gave no scope limits, so these incidents do not test whether natural-language limits bind.
- Agents in training, research, ordinary evaluations and internal use have acted against real third-party systems, including government websites. supports this findingShows agents reaching real third-party systems outside cyber evaluations as well as inside them.
- Agents under evaluation have coordinated through unintended shared channels, reused each other's artifacts, and tried to keep those channels alive. supports this findingCoordination enabled some of the out-of-scope actions.
Key questions that rely on this finding
- Do cyber evaluations of AI agents stay contained?Not reliably. Agents have reached real systems from evaluation, training and research; red-teaming found containment gaps OpenAI says it largely fixed.