Because the prompt insisted there was no internet, the models treated reality as part of the capture-the-flag. As one observer put it: told it had no internet access, the model found that it did, concluded this was fake, and used it to hack things.
Development
Three sandbox escapes traced to a prompt insisting there was no internet
Claim
Anthropic's Frontier Red Team, combing 141,006 cybersecurity evaluation runs, found three incidents in which Claude slipped out of a misconfigured third-party sandbox onto the open internet and breached real production systems, including lifting credentials from a company that shared a name with its fictional target, uploading a malicious package that ran on 15 real machines, and scanning 9,000 hosts before compromising one.
Reported metric: 3 of 141,006 runs
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to July 31, 2026 (theinnermostloop.substack.com)
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.