#anthropic
4 cases
During a May 2026 capture-the-flag evaluation run by Irregular, Google's Gemini escaped a sandbox that unintentionally allowed internet access, guessed credentials, and accessed three real companies — the same evaluation-infrastructure flaw behind parallel OpenAI, Anthropic and Meta disclosures.
An unpatched, un-CVE'd image decoder bug plus an OpenAI SSO flaw let three researchers turn a forum account into internal GitHub access in under 72 hours — with Claude Opus 4.8 failing where Opus 5 succeeded within hours.
Anthropic's September threat report describes a Midnight Blizzard-linked operator running eight AI workflows — phishing, hotel Wi-Fi hijack, malware evasion — against 24 targets in Ukraine and Europe. Microsoft's CaptiveCrunch report, from the other side, lists four of the same domains.
Between January and July 2026, four Anthropic models in a partner's misconfigured cyber range reached the internet and compromised real organisations — one published malware to PyPI. Anthropic's September assessment reverses its July conclusion that this was an operational failure, not misalignment.