Meta has confirmed that one of its AI models exploited a vulnerability in a third-party service during testing, joining OpenAI and Anthropic in reporting similar incidents.

Meta said the incident occurred during testing by independent firm Irregular.

A misconfiguration by Irregular allowed one of Meta’s models to access the internet during evaluation. The AI then went on to exploit a security vulnerability in a third-party service.

The tech giant said the exploit occurred in a manner similar to previously-reported instances with other companies.

“Meta learned of this when Irregular notified us, and we are currently investigating and will issue a full retrospective once we have all the facts," said a statement by a Meta spokesperson sent to Infosecurity.

Separately, on August 4, Open AI published an update describing how two external testing partners identified incidents in which testing configurations and controls allowed model activity to extend beyond the test environment.

One involved Irregular. OpenAI said that during a Capture-the-Flag-style evaluation intended to be isolated from the internet, a testing-environment misconfiguration allowed models to access the public internet.

The second related to the UK’s AI Security Institute (AISI), which said it had detected unusual data transfers leaving its research systems during a routine cyber evaluation.

This and the recent breaches by OpenAI and Anthropic models have raised security concerns among the cybersecurity community.

“When several of the world’s most capable AI systems reach real people, services and companies from test environments within a matter of weeks, we can no longer dismiss these as isolated incidents,” said Tim Hudson, president of OpenSSL, a free open-source software library that provides secure digital communications.

“We are seeing a recurring pattern: autonomous systems are given an objective, internet access and excessive authority - and those responsible only discover afterwards what the systems have done.”

This does not mean that AI has developed malicious intent of its own, rather poorly constrained objectives, vulnerable interfaces and permissions to act are allowing these acts to be carried out.

“We need to be careful to not assume that an AI independently decided to become a cybercriminal. It was given internet access, tools and an objective by people. The concern is that it was then able to chain actions together in ways its creators did not fully anticipate,” noted Javvad Malik, Lead CISO Advisor at KnowBe4.

AI Firms Criticized for One-Upmanship

Across the cybersecurity sector, many are becoming increasingly skeptical of the competition among vendors who claim their models are the most powerful.

“There also appears to be an underlying game of one-upmanship between AI vendors touting how powerful their models are,” Malik said.

Meanwhile, Alex Goller, Principal Solution Architect EMEA at Illumio, said that the fact that there has been three similar incidents across the biggest AI players is “simply ridiculous.”

“We've seen guardrails intentionally loosened to test their limits – Meta's model didn't need to be clever to breach another company's systems. The timing of conveniently finding the exact same problem either means it's a stunt or they weren't paying enough attention during testing. Either way, both answers are worrying,” he said.

AI Governance is Key

The key theme across the incident is that AI was given a task to conduct but was not constrained by sufficient guardrails to stop it behaving in a way that compromised external organizations.

Jack Nelson, CISO at Ivanti, said, "Security teams and their organizations need to carefully map a governance plan and policies for AI agents. As they become more powerful, so will their chances of conducting rogue activities that can have significant long-term impact.

Malik said the key takeaway from these issues is governance.

“As organizations give AI greater access to systems and data, human oversight, least-privilege permissions and effective monitoring become essential,” he said.

Robust guardrails and visibility into what actions AI is undertaking will need to be prioritized and invested in.

AI agents should be governed by least-privilege access, privacy-by-design principles and real-time monitoring.

Infosecurity has contacted Irregular for comment.

Image credit: Poetra.RH / Shutterstock.com