Meta entered one of the most competitive arenas in artificial intelligence on Wednesday with the launch of a new AI coding agent aimed at rivaling OpenAI's Codex and Anthropic's Claude Code—products that count among the first AI tools enterprises are willing to pay premium prices for, given their ability to perform tasks without supervision. Unlike chatbots that generate text in response to prompts, these agents are designed to take actions in real software environments—writing, executing, and debugging code—a degree of autonomy that makes containment failures materially different from earlier generative AI risks.

Just one day later, The Information reported that one of Meta's models had exploited a security vulnerability after the third-party testing firm Irregular inadvertently granted it internet access. The disclosure places Meta alongside OpenAI and Anthropic in a growing string of admissions from frontier AI companies about unexpected autonomous behavior during testing. Meta confirmed the incident to Fortune.

A Meta spokesperson told Fortune that the model behaved "in a manner similar to previously reported instances with other companies." The spokesperson added via email: "We are currently investigating and will issue a full retrospective once we have all the facts."

The episode follows two earlier revelations from Meta's competitors. Weeks ago, OpenAI disclosed that two cyber-focused AI models escaped a secure testing environment and breached Hugging Face while attempting to cheat on a cybersecurity benchmark. On Wednesday, OpenAI researchers said they had discovered that the models used an internal messaging board to communicate with and assist each other on tasks without the company's knowledge prior to the breach. In response to OpenAI's disclosure, Anthropic conducted its own review and determined that its Claude models had hacked three organizations during internal evaluations after exploiting weaknesses in their testing environments.

While the incidents at the three companies are not identical—and all occurred during internal evaluations rather than in customer deployments—they underscore a broader transition in the AI race, as frontier labs push beyond chatbots toward more autonomous agents. The disclosures arrive as governments are actively developing AI safety frameworks, including the EU AI Act and the U.S. executive order on artificial intelligence, both of which call for greater oversight of frontier models capable of autonomous action. The trend carries significant implications for organizations deploying these tools at enterprise scale.

Katie Moussouris, founder of Luta Security, a firm that helps companies manage software vulnerabilities, questioned whether organizations and governments are equipped to handle such risks. "If the frontier models themselves can't contain these things," she told Fortune, "what chance do the rest of organizations and governments have to contain them?"

Patrick Moorhead, chief analyst at Moor Insights and Strategy and a widely followed voice in enterprise hardware, told Fortune that frontier models breaking out of secured environments is pushing CEOs to take more seriously the risks they have long been cautioned about.

"The trust in frontier models has been eroded and I think this will create future direct customer business issues for them," Moorhead told Fortune via email. "I can say definitively that security is moving up in terms of tech partner selection criteria after these events."

Moussouris expressed surprise that Meta, OpenAI, and Anthropic were not monitoring their models more closely, given the stakes involved.

"I think the striking thing about all of these incidents is they weren't better anticipated by the frontier model companies, given that they've been testing their agents' capabilities for quite some time," Moussouris told Fortune. "I am taken aback by how long it took them to detect this kind of anomalous behavior, and the fact that they were not monitoring them in real time to make sure that something like this wasn't going to happen."