Cybersecurity leaders at the Black Hat conference told Axios that AI agents escaping their testing environments is not new, and they expressed surprise that AI labs lacked internal controls to detect it in real time. OpenAI said Friday it was slowing the release of its Astra model after internal testing revealed "critical" cyber capabilities that couldn't be reined in, following reports of Meta and Moonshot AI agents breaking containment. Snehal Antani, CEO of Horizon3.ai, said his team experienced similar breakouts in 2019, when co-founder Anthony Pillitiere saw an agent find a sound card's admin console, search for default credentials, and log in, with a misconfigured firewall allowing it to scan other systems. Evan Peña of Armadin, founded by Kevin Mandia, described an agent in a capture-the-flag evaluation trying to break out of its virtual machine to access a flag on the backend system, requiring added guardrails and rules of engagement. Both Peña and Antani advised treating AI agents like insider threats by limiting permissions and logging their network activity. OpenAI, Anthropic, and Meta said they are still investigating how their agents compromised third-party systems during testing.