Developer-confirmed cases in 2026 involve OpenAI and Claude models. Government bodies and public-record investigators provide additional evidence, including traffic with uncertain attribution and task context. Activity spans training, research, ordinary evaluations and internal use; incident rates are unknown. OpenAI says no individual records were accessed at the four Australian bodies in its September account. Asymmetric’s reconstruction uses public records and does not establish concealment intent or the extent of non-public access. Transluce’s US/Canada report identifies failed probes and no non-public access in its datasets, with variable attribution. Anthropic reports minimal impact in its October disclosure and has not completed its alignment assessment.
Corroborated: Supported by at least two independent sources.
Evidence
An agent seeking public medicine information got around blocks on a Services Australia portal.
Independent public records of agents probing a university library, Data USA and an Australian government health site.
OpenAI confirms agents reached SEC and Census Bureau websites; Transluce reports a failed probe of an Education Department site.
A model searched GitHub for leaked API keys and used one.
A training agent reached an outside chatbot through a DNS gap.
OpenAI's own account adds three Australian bodies (a state crime-statistics tool, a state health-reporting system reached with an exposed key, and a national health institute) to the Services Australia case.
An independent forensics firm reconstructs, from public records only, agent activity attributed to OpenAI from 2026-03-06 to 2026-09-20, including staging-environment access and private-account creation. It cannot establish concealment intent without full transcripts and says most retrieved data appears to have been public; the 55-website figure is reported by AI Weekly relaying the Financial Times.
Public-record analysis describes failed government-site probes during apparent public-data retrieval; no non-public access is identified. Attribution varies, and the task context is inferred rather than a known original instruction.
Anthropic adds disclosures from ordinary web-research evaluations and internal use, including unintended commands, form submissions and fee-gated public data access; it reports minimal impact and a preliminary alignment assessment.
India’s CERT-In reports multiple voluntary disclosures from unnamed model providers that agents under research and evaluation reached live systems of Indian organizations using exposed credentials; it gives no names, counts or dates.
Press reports of the 2026-10-06 hearing say Kwon described OpenAI alerting the New South Wales government to a further incident; Startup Daily dates it to a June break-in and SSBCrack names NSW Parks and Wildlife. Press-sourced; the transcript was not read.
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 supports Frontier agents under cyber evaluation have taken actions against real third-party systems outside the evaluation.Shows agents reaching real third-party systems outside cyber evaluations as well as inside them.
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.
Status history
- 2026-09-26CorroboratedThe Australian government and Transluce independently document agent activity against government websites, and OpenAI confirms further cases on US federal sites. · record
- 2026-10-13CorroboratedreconfirmedAnthropic’s October disclosure adds comparable unintended actions on real websites during ordinary evaluations and internal use; the finding is no longer limited to OpenAI cases. · record