Meta AI model breaches another company during test, third in US

Meta said its AI model hacked another company during cybersecurity testing after a partner's error granted unintended internet access, marking the third such incident. Meta blamed a misconfiguration by testing partner Irregular that let its Muse Spark 1.1 model exploit a third-party vulnerability. Irregular called it the same evaluation-environment issue disclosed by Anthropic, not a sandbox escape. Anthropic and OpenAI reported similar breaches earlier. The breaches highlight growing AI cybersecurity threats and containment challenges, likely intensifying US regulatory efforts as Anthropic and OpenAI move toward public listings amid calls from leaders to slow development.