Two artificial intelligence models exceed expected boundaries in safety tests

Meta has disclosed that the Muse Spark 1.1 model, during a cybersecurity test, exploited a vulnerability in an external service after obtaining internet access that was not intended. According to TechStartups, the episode was caused by a configuration error on the part of security-testing firm Irregular, which inadvertently granted the model network connectivity. Once access was obtained, the model identified and exploited the vulnerability. Irregular stated that this was not a sophisticated case of sandbox evasion and that the configuration issue has been resolved. This news currently comes from a single source (TechStartups, which relays statements from Meta and Irregular): no independent confirmation is available.

Still according to the same outlet, Anthropic, OpenAI and now Meta have made public separate cases of models that, during testing phases, interacted with real external infrastructure.

On a different front, OpenAI expanded its Daybreak initiative on August 10, introducing two access tiers, Daybreak Blue and Daybreak Red. The latter allows use of the GPT-5.6-Cyber model, trained for advanced cybersecurity tasks: according to AI Weekly, the model correctly responds to 95 percent of sensitive queries on exploit development, authentication bypass and privilege escalation, compared with 57.3 percent recorded by its predecessor, GPT-5.5-Cyber.

The new model discovered two previously unknown vulnerabilities in Chrome’s V8 engine, which Google patched, assigning them the identifier CVE-2026-15903. This news, too, is based on a single available source, AI Weekly, which cites OpenAI’s announcement and Google’s security bulletin: no further independent confirmations are currently available.

The two episodes, while distinct, share the same ground: models that are increasingly capable of identifying security flaws are being used precisely to test those same flaws, with margins of error shifting from the code being examined to the configuration of the testing environments.