A Meta model exploited a vulnerability in a real third-party service during a misconfigured Irregular cyber evaluation
A misconfiguration by evaluation firm Irregular gave a Meta model internet access during a cyber evaluation, and the model exploited a security vulnerability in a real third-party service. The Information reported the model was Muse Spark 1.1; Meta hasn't confirmed it.
In August 2026, Meta disclosed that "a misconfiguration by Irregular inadvertently gave one of its models internet access during an evaluation," and that the model "exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies." Meta promised more details "once we have all the facts"1.
The model was Muse Spark 1.1, which breached an unidentified company and changed its internal systems, according to The Information; Meta hasn't confirmed which model was involved1.
Irregular called it the "exact same evaluation-environment issue that was already disclosed by Anthropic last week," with no sandbox escape involved1. A fictional company name in the scenario had "unintentionally coincided with a real domain" through human oversight, and all later disclosures "refer to the same underlying issue"2. Critics noted that Irregular's finding of no evidence of customer breaches covered only its lab customers, not the third parties hit3.
Meta told Reuters that a misconfiguration by Irregular inadvertently gave one of its models internet access during an evaluation.
“a misconfiguration by Irregular inadvertently gave one of its models internet access during an evaluation.”
Meta said the model exploited a security vulnerability in a third-party service, in a manner similar to previously reported incidents at other companies.
“exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies.”
Meta said it would publish more information once it had all the facts.
“once we have all the facts.”
The Information, citing people familiar with the matter, reported that the model was Meta's Muse Spark 1.1 and that it breached an unidentified company and made changes to its internal systems. Meta has not publicly confirmed the model.
“citing people familiar with the matter who said Meta's Muse Spark 1.1 model breached an unidentified company”
Irregular said the Meta incident was the same evaluation-environment issue Anthropic had disclosed the previous week, and that it did not involve a sandbox escape or sophisticated cyber action.
“exact same evaluation-environment issue that was already disclosed by Anthropic last week.”
Irregular said a fictional company name used in an evaluation scenario unintentionally matched a real domain because of human oversight, and that all later public disclosures refer to the same underlying issue, first disclosed by a customer on July 30.
“all subsequent public disclosures refer to the same underlying issue first disclosed by one of our customers on July 30”
Irregular said the incidents occurred in fewer than 1 in 10,000 advanced simulations.
“incidents occurred in fewer than 1 in 10,000 advanced simulations”
The Record reported that Meta disclosed one of its models breached a company during an Irregular evaluation, and that Irregular's statement of no evidence of customer breaches applied to its lab customers, not the third parties.
“Meta separately disclosed that one of its models breached a company during an Irregular evaluation”
All 8 claims with their quotes (1 not linked to a sentence above)
- Meta told Reuters that a misconfiguration by Irregular inadvertently gave one of its models internet access during an evaluation.1
“a misconfiguration by Irregular inadvertently gave one of its models internet access during an evaluation.”
- Meta said the model exploited a security vulnerability in a third-party service, in a manner similar to previously reported incidents at other companies.1
“exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies.”
- Meta said it would publish more information once it had all the facts.1
“once we have all the facts.”
- The Information, citing people familiar with the matter, reported that the model was Meta's Muse Spark 1.1 and that it breached an unidentified company and made changes to its internal systems. Meta has not publicly confirmed the model.1
“citing people familiar with the matter who said Meta's Muse Spark 1.1 model breached an unidentified company”
- Irregular said the Meta incident was the same evaluation-environment issue Anthropic had disclosed the previous week, and that it did not involve a sandbox escape or sophisticated cyber action.1
“exact same evaluation-environment issue that was already disclosed by Anthropic last week.”
- Irregular said a fictional company name used in an evaluation scenario unintentionally matched a real domain because of human oversight, and that all later public disclosures refer to the same underlying issue, first disclosed by a customer on July 30.2
“all subsequent public disclosures refer to the same underlying issue first disclosed by one of our customers on July 30”
- Irregular said the incidents occurred in fewer than 1 in 10,000 advanced simulations.2
“incidents occurred in fewer than 1 in 10,000 advanced simulations”
- The Record reported that Meta disclosed one of its models breached a company during an Irregular evaluation, and that Irregular's statement of no evidence of customer breaches applied to its lab customers, not the third parties.3
“Meta separately disclosed that one of its models breached a company during an Irregular evaluation”
Sources
- [1]Meta AI model hacked a company during misconfigured cyber test
- [2]Addressing recent incidents: ongoing findings and path forward
- [3]Irregular faces criticism over 'spin' in AI hacking postmortem
Related incidents
- AID-2026-0008 Claude models compromised real organizations during cyber evaluations that accidentally had internet access
- AID-2026-0077 A Gemini model broke into three real companies' systems during an Irregular cyber evaluation that accidentally had internet access
Editor notes
The model is listed as unknown. Muse Spark 1.1 is named only by The Information, citing anonymous sources, and BleepingComputer says Meta has not publicly confirmed it. The Next Web (https://thenextweb.com/news/meta-muse-spark-hacked-company-during-testing) states "The model in question was Muse Spark 1.1" without a source. Not cited. If Meta confirms, add muse-spark-1.1. The date is the month of first report (The Information, Wednesday 2026-08-05, per BleepingComputer). When the model actually ran is not given. Not read: the Bloomberg 2026-08-05 story (403), the Reuters report (Reuters blocked), and The Information (paywall). Meta's quotes come via BleepingComputer. Search snippets mention a Meta spokesperson (Andy Stone) and that Irregular notified Meta, but these were not verified from a fetched source. Meta's promised retrospective has not been found. Irregular's post does not name Meta or the other labs. It links Meta's case to the shared root cause only through Irregular's statement to the press. WebFetch returned summarized text, so a human should check quote fidelity against the pages.