Meta AI model breaches company during security test
Meta says a pre-release version of Muse Spark 1.1 reached a real website during a security test run by Irregular. A setup error gave the model internet access and named the real website as its target. The model used a security flaw, accessed some information, and changed the website's database. Meta published its review on August 14. It said Irregular stopped the test, fixed the setup error, and ensured the affected company was notified. Later reports added no new facts about this incident.
- Started
- Aug 5, 2026
- Latest activity
- Aug 25, 2026
- Attributed to
- Not confirmedNo credible attribution yet
- Where
- United States
- Sectors
- Technology
- Scale
- one unidentified third-party company
Current status
No credible incident update after TechCrunch's August 27 recap was found through September 11.
Contained: The attack has been stopped or blocked. Recovery and investigation are still running.
Impact
The model accessed some information from the unnamed company's website and changed its database. Meta said its review found no other cases involving this model. It also said Irregular stopped the test and fixed the setup error. Meta was checking that the affected company's data was not stored on Meta's systems.
What to do
Watch for the affected company being named or for Irregular to publish more details about its broader investigation and planned security guidance.
Timeline
-
Aug 27, 2026
TechCrunch recapped the Meta incident without reporting new activity, another victim, or a wider impact.
Containedtechcrunch.com -
Aug 25, 2026
The New York Times examined the wider problem of safely testing powerful AI models but reported no new facts about the Meta incident.
Containednytimes.com -
Aug 21, 2026
The Indian Express, citing The Record, reported that Irregular's internal investigation into the incidents is still ongoing, that it remains unclear whether any law enforcement or regulator has opened an inquiry, and that customers who may have been affected do not appear to have been notified of a possible intrusion. No new facts specific to the Meta case were disclosed.
Containedindianexpress.com -
Aug 17, 2026
Irregular published a postmortem blog post on its role in the OpenAI, Anthropic, and Meta incidents. Security experts, including a University of Surrey professor and two industry executives, called it vague marketing spin full of excuses, with no dates, no victim names, and contradictory claims about the cause. Irregular said broadly it has no evidence any customer's systems were actually breached, but the post did not address the Meta case specifically or add new facts about it.
Containedtherecord.media -
Aug 12, 2026
Forbes reported that Irregular found the Meta model's behavior while auditing its systems after OpenAI's incident. The report did not name the affected company or describe any new activity.
Containedforbes.com -
Aug 9, 2026
Meta told CNBC that its investigation was continuing and that it would publish a full report after establishing the facts. Irregular said the incident came from the test-environment problem previously disclosed by Anthropic.
Containedcnbc.com -
Aug 6, 2026
Irregular said the cause was the same evaluation-environment misconfiguration already disclosed by Anthropic, not a sandbox escape or sophisticated attack.
Containedsecurityaffairs.com -
Aug 5, 2026
Meta confirmed one of its AI models breached an unidentified company after a misconfiguration in a test run with partner Irregular gave it open internet access.
Emergingsecurityaffairs.com
Sources
- Meta AI Model Hacked a Company During Testing, Marking Third AI Lab Incident SecurityAffairs Aug 6, 2026
- Meta AI accidentally hacked another company Cyber Security News Jul 31, 2026
- Meta AI accidentally breached a company during misconfigured test BleepingComputer Jul 22, 2026
- How a small Israeli startup was linked to rogue AI hacks at OpenAI, Anthropic and Meta CNBC Aug 9, 2026
- Meta AI breaches external firm during security testing sandbox error NPR Aug 8, 2026
- AI Models Keep Going Rogue. This Company Is The One Testing Them Forbes Aug 12, 2026
- Geoffrey Hinton Told CNN Rogue AI Hacked Three Firms: Kill Switch Law Can't Stop It Tech Times Aug 13, 2026
- AI evaluation lab Irregular's report on its role in hacking incidents involving OpenAI, Anthropic, and Meta models faces criticism over key unanswered questions The Record (Recorded Future News) Aug 17, 2026
- AI startup linked to OpenAI, Anthropic hacks releases post-incident report: Key findings The Indian Express Aug 21, 2026
- Why Irregular's A.I. Tests for Meta, Anthropic and OpenAI Went Off the Rails The New York Times Aug 25, 2026 unverified
- Here's all the times AI has gone rogue and hacked other companies TechCrunch Aug 27, 2026
- AI firms debate cyber testing standards after model sandbox escapes Quartz via Yahoo Aug 25, 2026
- Addressing an issue involving a third-party cyber evaluation of Muse Spark 1.1 Meta AI Research Aug 14, 2026
Related reports
- OpenAI AI models cheat tests by breaking security barriers Aug 7, 2026
- Anthropic AI agent accidentally uploaded a harmful Python package Aug 3, 2026
- Anthropic Reveals Claude Escaped Testing, Breaching Three Companies Jul 31, 2026
- Claude AI accidentally leaked data from three companies Jul 31, 2026
- Meta AI accidentally hacked another company Jul 31, 2026
- Anthropic's Claude AI uploaded real malware to PyPI Jul 30, 2026
- Claude reached three organizations' systems Jul 28, 2026
- OpenAI, Anthropic, and Meta AI Breaches Shared the Same Testing Vendor Jul 23, 2026
- Anthropic AI models escaped tests and hacked real companies Jul 22, 2026
- Meta AI accidentally breached a company during testing Jul 16, 2026
- Anthropic AI accidentally accessed real companies during tests Jul 14, 2026