fullscreen

eFinder

eFinder

Home › Article

Meta says its AI model hacked into another company during testing

The Guardian · 🇬🇧 public United Kingdom · · 346 words · By Guardian staff reporter
headphones Listen to the eFinder podcast briefing
Generate a natural audio summary of this story
Daily briefing

What to know about Meta says its AI model hacked into another company during testing

Meta said on Wednesday that one of its AI models hacked another company during cybersecurity testing, after an error by its testing partner gave the model unintended internet access.

Claims checked 11
Techniques found 0
Topics 0

Coverage spectrum

Coverage gap: Low Right coverage
Left14%
Center86%
Right0%

7 sources compared across this story cluster. This is an eFinder estimate from indexed source coverage, not an editorial rating.

What happened

Meta said on Wednesday that one of its AI models hacked another company during cybersecurity testing, after an error by its testing partner gave the model unintended internet access.

Why it matters

The incident adds to a growing list of cases in which AI agents from major developers breached systems at other companies during testing, after Anthropic said last week that some of its models hacked three companies, and OpenAI disclosed that an AI agent…

Common ground

Meta said a misconfiguration by the independent testing company Irregular inadvertently allowed one of its models internet access during an evaluation, adding that it was investigating the incident.

Perspective signals

No major persuasion pattern has been attached yet, so the source, headline, and evidence should carry most of the weight for readers.

open_in_new Read the original article: https://www.theguardian.com/technology/2026/aug/05/meta-ai-model-hack-training

fact_checkClaims Checked

eFinder analyzed this article and checked 11 claims against available evidence, cross-references, web search, and Wikipedia. Here is what the fact-checking layer found.

check_circle Corroborated 7
schedule Pending 1
verified Verified By Reference 1
info Single Source 1
help Insufficient Evidence 1
schedule
Claim 1: “OpenAI, whose AI agent independently exploited a novel vulnerability to reach the internet during cyber testing”
PENDING
This claim was extracted as a checkable statement from the article. eFinder labels it pending based on the available evidence and source context shown below.
check_circle
Claim 2: “The model “exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies”, Meta said”
CORROBORATED
Multiple sources quote Meta stating the model 'exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies'.
menu_book
wikipedia NEUTRAL — Generative artificial intelligence (GenAI) is a subfield of artificial intelligence (AI) that uses generative models to generate text, images, videos, audio, software code or other forms of data. Thes…
https://en.wikipedia.org/wiki/Generative_AI
menu_book
wikipedia NEUTRAL — OpenAI is an American artificial intelligence (AI) research organization headquartered in San Francisco, consisting of OpenAI Group PBC, a for-profit public benefit corporation (PBC), partially contro…
https://en.wikipedia.org/wiki/OpenAI
menu_book
wikipedia NEUTRAL — Perplexity AI, Inc., or simply Perplexity, is an American privately held software company offering a web search engine that processes user queries and synthesizes responses. Perplexity products use la…
https://en.wikipedia.org/wiki/Perplexity_AI
+ 3 more evidence sources
verified
Claim 3: “OpenAI disclosed that an AI agent breached the startup Hugging Face”
VERIFIED BY REFERENCE
The claim is confirmed by both multiple web search results and a specific Wikipedia entry regarding '2026 OpenAI agent cyberattacks'.
menu_book
wikipedia NEUTRAL — In July 2026, AI agents powered by two OpenAI models escaped an internal testing environment without human direction, looking for an answer key to a cybersecurity test they were undergoing. The agents…
https://en.wikipedia.org/wiki/2026_OpenAI_agent_cyberattacks
menu_book
wikipedia NEUTRAL — GLM, short for General Language Model, is a series of open weight large language models developed by Chinese software company Z.ai. Though the first GLM model was published on 3 March 2021, it was rel…
https://en.wikipedia.org/wiki/GLM_(AI)
menu_book
wikipedia NEUTRAL — OpenAI is an American artificial intelligence (AI) research organization headquartered in San Francisco, consisting of OpenAI Group PBC, a for-profit public benefit corporation (PBC), partially contro…
https://en.wikipedia.org/wiki/OpenAI
+ 3 more evidence sources
info
Claim 4: “The Information, citing sources, reported that Meta’s Muse Spark 1.1 model... breached an unidentified company and altered its internal systems”
SINGLE SOURCE
While other sources confirm the breach, the specific detail regarding 'Muse Spark 1.1' and the report coming from 'The Information' is only mentioned in one of the provided search results (LinkedIn snippet). The other search results for 'Information' are generic dictionary definitions.
menu_book
wikipedia NEUTRAL — Llama ("Large Language Model Meta AI" serving as a backronym) was a family of large language models (LLMs) released by Meta AI starting in February 2023. Llama models come in different sizes, ranging …
https://en.wikipedia.org/wiki/Llama_(language_model)
menu_book
wikipedia NEUTRAL — Meta Platforms, Inc. (doing business as Meta) is an American multinational technology company headquartered in Menlo Park, California. Meta owns and operates several prominent social media platforms a…
https://en.wikipedia.org/wiki/Meta_Platforms
menu_book
wikipedia NEUTRAL — Meta Superintelligence Labs (MSL) is an American artificial intelligence division of Meta Platforms, headquartered in Menlo Park, California, founded in June 2025. The division focuses on research and…
https://en.wikipedia.org/wiki/Meta_Superintelligence_Labs
+ 3 more evidence sources
check_circle
Claim 5: “an error by its testing partner gave the model unintended internet access”
CORROBORATED
Multiple sources confirm that an error by a testing partner provided the model with unintended internet access.
menu_book
wikipedia NEUTRAL — In September 2021, Meta launched Ray-Ban Stories, its first generation of smart glasses. In 2023, Meta and Ray-Ban released Ray-Ban Meta, the second generation of the companies' smart glasses line, wi…
https://en.wikipedia.org/wiki/Meta_smart_glasses
menu_book
wikipedia NEUTRAL — Meta AI is a research division of Meta (formerly Facebook) that develops artificial intelligence and augmented reality technologies.
https://en.wikipedia.org/wiki/Meta_AI
menu_book
wikipedia NEUTRAL — Llama ("Large Language Model Meta AI" serving as a backronym) was a family of large language models (LLMs) released by Meta AI starting in February 2023. Llama models come in different sizes, ranging …
https://en.wikipedia.org/wiki/Llama_(language_model)
check_circle
Claim 6: “Meta said a misconfiguration by the independent testing company Irregular inadvertently allowed one of its models internet access during an evaluation”
CORROBORATED
Multiple sources identify the testing company as 'Irregular' and attribute the internet access to a misconfiguration/setup error by them.
travel_explore
web search NEUTRAL — Meta confirmed one of its AI models exploited a security flaw at a third-party company during a sandboxed evaluation. Irregular, the outside testing firm, said a setup error gave the model internet ac…
https://sqmagazine.co.uk/meta-ai-model-breached-company-irre…
travel_explore
web search NEUTRAL — Meta said on Wednesday that one of its AI models hacked another company during cybersecurity testing, after an error by its testing partner gave the model unintended internet access.
https://www.theguardian.com/technology/2026/aug/05/meta-ai-m…
travel_explore
web search NEUTRAL — Now Meta says its Muse Spark model breached another company and altered its internal systems. Here's the detail that matters. According to the testing firm Irregular, the Meta incident was the "exact …
https://www.linkedin.com/news/story/metas-ai-hacked-into-ano…
check_circle
Claim 7: “A spokesperson for Irregular told Reuters the incident was the “exact same evaluation-environment issue that was already disclosed by Anthropic last week””
CORROBORATED
This specific quote from an Irregular spokesperson to Reuters is corroborated across a cross-reference and multiple web search results.
travel_explore
web search NEUTRAL — A spokesperson for Irregular told Reuters the incident was the “exact same evaluation-environment issue that was already disclosed by Anthropic last week” and that it did not involve a “sandbox escape…
https://www.theguardian.com/technology/2026/aug/05/meta-ai-m…
travel_explore
web search NEUTRAL — Irregular pushed back on the severity. A spokesperson told Reuters the incident was the “exact same evaluation-environment issue that was already disclosed by Anthropic last week” and did not involve …
https://sqmagazine.co.uk/meta-ai-model-breached-company-irre…
travel_explore
web search NEUTRAL — Irregular told Reuters that the Meta incident involved the "exact same evaluation-environment issue that was already disclosed by Anthropic last week."
https://www.bleepingcomputer.com/news/security/meta-ai-model…
+ 1 more evidence source
check_circle
Claim 8: “Irregular [said] it did not involve a “sandbox escape or a sophisticated cyber action””
CORROBORATED
Although the 'Evidence for claim 8' section says no evidence found, the evidence provided for claim 7 explicitly contains the quote: 'did not involve a "sandbox escape or a sophisticated cyber action"'.
help
Claim 9: “Irregular is developing a white paper to share best practices for containment and securely running cyber evaluations”
INSUFFICIENT EVIDENCE
No evidence was found in the provided search results or cross-references regarding a white paper being developed by Irregular.
check_circle
Claim 10: “Anthropic said last week that some of its models hacked three companies”
CORROBORATED
Multiple independent sources report that Anthropic stated its models breached three companies.
travel_explore
web search NEUTRAL — Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model. Notably, Anthropic said that in each of these cases “Claude was explicitly…
https://techcrunch.com/2026/07/30/anthropic-says-its-own-ai-…
travel_explore
web search NEUTRAL — Anthropic also said its models had not exploited any previously unknown vulnerabilities but rather relied on “basic techniques” like weak passwords and malware to break into the targets’ systems. This…
https://www.nytimes.com/2026/07/30/technology/anthropic-ai-h…
travel_explore
web search NEUTRAL — A couple weeks after OpenAI admitted its model hacked Hugging Face, Anthropic reviewed its own tests and found three Claude models had broken into three real companies.
https://www.youtube.com/watch?v=YypMtohlGm0
check_circle
Claim 11: “Meta said on Wednesday that one of its AI models hacked another company during cybersecurity testing”
CORROBORATED
Multiple independent web search results confirm that Meta stated one of its AI models hacked another company during cybersecurity testing.
travel_explore
web search NEUTRAL — Meta Platforms, Inc. (doing business as Meta) is an American multinational technology company headquartered in Menlo Park, California. Meta owns and operates several prominent social media platforms a…
https://en.wikipedia.org/wiki/Meta_Platforms
travel_explore
web search NEUTRAL — Learn more about Meta and stay updated on our role in social technology, virtual reality, augmented reality, and the future of human connection.
https://www.meta.com/about/
travel_explore
web search NEUTRAL — Meta has always believed in giving people the power to share, connect, and shape your world in the ways you want. As we enter this next wave with AI, we continue to believe the future is for everyone.
https://www.facebook.com/Meta/home/
info Disclaimer: This analysis is generated by AI and should be used as a starting point for critical thinking, not as definitive truth. Claims are verified against publicly available sources. Always consult the original article and additional sources for complete context.