Evidence · DiggingBeagle record
Investigating three real-world incidents in our cybersecurity evaluations
Anthropic's initial disclosure describes three cybersecurity evaluation incidents across six runs, found in a review of 141,006 runs. The models reached real organizations through unintended internet access in an external evaluation environment.
- Published
- Jul 30, 2026
- Accessed
- Sep 14, 2026
- Publisher
- Anthropic
- Source role
- primary
- Version
- 2026-07-30
- Rights
- Public web source; citation and short excerpt only.
Evidence record
Anthropic reviewed 141,006 evaluation runs and reported three incidents involving unauthorized access to real organizations.
Claim-level citations (4)
- supportsClaude cyber evaluations reached real third-party systems: In four Opus 4.7 runs against the same real company, Anthropic says the model obtained credentials and accessed a database containing several hundred rows.
first incident description
- supportsClaude cyber evaluations reached real third-party systems: Anthropic says an internal research model scanned roughly 9,000 targets, compromised an internet-facing application, then stopped after recognizing that the system appeared real.
third incident description
- contextClaude cyber evaluations reached real third-party systems: Anthropic revised its earlier interpretation and concluded that the incidents showed biased reasoning and recklessness in addition to infrastructure failure.
Initial disclosure, interpretation
- supportsClaude cyber evaluations reached real third-party systems: Anthropic says the evaluation infrastructure was segregated from sensitive internal systems and customer data, so the incidents did not expose Anthropic customer data.
infrastructure separation and impact discussion