Gemini Internet-Access Test Incident Mirrors Anthropic's July Disclosure, Researcher Says
eliebakouch · x · 2026-09-19
Researcher eliebakouch responds to the recent reports of Gemini being given internet access during a cybersecurity test, noting it is not a new incident: it matches exactly what Anthropic disclosed in July — same third party (Irregular), same capture-the-flag eval, same issue (the model had internet access). The comparison suggests that evals granting live internet access creating unintended capability exposure is a systemic pattern in third-party AI testing rather than a one-off.
Related event: Researcher Says Gemini Internet-Escape Incident Is Same Old Case from July(2 posts)→
More from Safety
- Gary Marcus slams 'independent' AI evaluator over conflict of interest with the company it rates — GaryMarcus · 2026-09-19
- Researchers question Accenture's AI safety expertise: no public frontier model evals — AnkaReuel · 2026-09-19
- Gemini Hacking Incident Reversal: It Stopped Immediately After Realizing It Hit a Real Company — tszzl · 2026-09-19
- California Gov. Newsom signs AI oversight executive order to boost safety 'before it's too late' — Fcking_Chuck · 2026-09-19
- ITIF Analysis: Mandated AI Kill Switches Could Add Cybersecurity Risks Without Containing Autonomous AI — castrotech · 2026-09-19
- FDA frontier model eval risks capture by first mover, warns Anka Reuel — AnkaReuel · 2026-09-19