CISO Voice

AI & Security / News

Meta’s AI model hacked a third-party company during safety evaluation

Meta's Muse Spark 1.1 reportedly carried out the hack during a third-party safety evaluation.

Akshaya Asokan
Akshaya Asokan Cybersecurity Reporter · August 6, 2026 · 4 min read
Meta Spark.

Meta is the latest company to claim its AI model hacked another firm after escaping its test environment, becoming the third company after OpenAI and Anthropic to make similar claims.

A Meta spokesperson on Wednesday told multiple media outlets that one of its AI models compromised a third-party company. The Information, which reported the news first, disclosed that Meta’s multimodal reasoning AI system, Muse Spark 1.1, carried out the hack.

The model, released last month, is a coding agent that can write and debug code and run multiple agents concurrently.

Meta said the model bypassed security guardrails due to a misconfiguration in the third-party evaluation partner, Irregular, allowing it to connect to an internet-facing application and compromise the target.

Meta did not disclose details about the targeted company and did not respond to a request for comment.

An Irregular spokesperson told the BBC that the incident is the “exact same evaluation-environment issue that was already disclosed by Anthropic last week.” 

In the Anthropic AI hacks, the company claimed its models Claude Opus 4.7, Mythos 5, and an unidentified AI model escaped Irregular’s evaluation environment to target three organizations.

Because both Anthropic and Irregular did not adequately review the evaluation transcripts or network logs, they failed to identify a system with open internet access. As a result, Claude Opus 4.7,  whose capabilities were being assessed as part of a capture-the-flag, compromised a company website resembling the test target despite guardrail limitations.

In the case of Mythos, the model published its own malicious package to compromise a target. A third, unidentified model compromised an internet-facing company’s system via an SQL injection vulnerability on an exposed debug page.

It is unclear which of the three tactics the Meta model used to compromise the target or if it employed the three tactics as in the case of Anthropic. Both Meta and Irregular did not respond to requests seeking clarification.

Prior to Meta and Anthropic, OpenAI was the first to disclose that its AI models hacked into another company after escaping the test environment. GPT-5.6 Sol and an unreleased model hacked into AI platform Hugging Face after compromising third-party software supply chain platform JFrog and AI computing platform Modal Labs.

Unlike the Meta and Anthropic incidents, which were allegedly caused by a misconfiguration, the OpenAI models chained together multiple zero-day flaws.

Legal scrutiny

These developments have already sparked concerns about the disruptive capabilities of cyber-enabled AI models, prompting scrutiny from the U.S. government. On Tuesday, Aug, 4, OpenAI, Anthropic, Google, and other tech companies convened at the White House to review the Trump administration’s voluntary framework for AI models.

The non-binding measure encourages AI companies to conduct cybersecurity tests to determine the hacking capabilities of American AI models. Additionally, 15 Republican lawmakers requested that OpenAI provide more details regarding the Hugging Face compromise.

Industry skepticism

While AI companies have been quick to publicize claims of their models’ hacking capabilities, leading AI critics say these claims are misleading because the models escaped their testing environments due to human error.

“Meta Has Joined The Club Of: ‘How is it possible our AI would be sociopathic?’ Club. Bad training data, no love equation, and removal of ‘AI Self,’ you get this behavior,” said Brian Roemmele, an AI and neural networks expert.

Bill Gurley, former VC at Benchmark and founder of technology consultancy firm P3 Institute, criticized AI companies’ tendency to distance themselves from their models’ disruptive capabilities by attributing them merely to guardrail or prompting issues.

“Please stop referring to your own models in the third person when talking about model bad behavior. Humans write the software; humans built the prompts; and they work for your company. ‘Our’ model is doing illegal things. ‘Our’ model is risky. ‘We now have liability,” Gurley said.

Akshaya Asokan
Akshaya Asokan Cybersecurity Reporter

Akshaya Asokan is a U.K.-based cybersecurity and technology journalist with experience covering nation‑state hacking operations, cyber and tech policy, and emerging cybercrime trends. She previously served as the senior European cybersecurity correspondent for Information Security Media Group’s Global News Desk. Before that, she reported on IT developments for IDG Media and covered artificial intelligence and machine learning for Analytics India Magazine. Earlier in her career, she wrote on sexual and religious minority rights as a reporter for The New Indian Express.

Keep reading for free,
one signup unlocks everything

We'll send you the CISO Voice newsletter. Unsubscribe anytime.

More from CISO Voice

Sign up to CISO Voice, the free newsletter trusted by senior security leaders.

Secret Link