News

Anthropic says its own AI models breached three companies during security tests

Published

4 months ago

July 31, 2026

Anthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.

Anthropic said the July 21 OpenAI incident prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did. Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Either way, because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. 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 told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.

That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.

The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models, safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the two incidents are inevitable, Anthropic drew a clear distinction between its incidents and OpenAI’s, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

OpenAI has continued to release new details about its own breach, saying its models also used publicly exposed credentials across four accounts on four services: one as a staging point, one for storage, and two that were only looked at, not used to break in further, according to OpenAI’s own updated blog post about the incident.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic.

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, sparked a string of reactions from the industry and politicians, many of whom don’t necessarily agree with one another. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Related Topics:Anthropic breached companies models security tests

Up Next

How Systems-Driven Education Interventions Can Rebuild Nigeria’s Workforce

Don't Miss

FG to offer 25,000 free cataract surgeries nationwide by 2027

Click to comment

News

Gov Alia receives APC flag for 2027 Benue governorship race

Published

49 minutes ago

November 15, 2026

PlateauReports

Benue State Governor, Rev. Fr. Hyacinth Alia, has received the All Progressives Congress, APC, flag as its governorship candidate for Benue State ahead of the 2027 election during the Tinubu/Shettima rally in Jos over the weekend.

The presentation of the flag by the APC National Chairman, Professor Nentawe Yilwatda, represents a formal affirmation of Governor Alia's recognition and acceptance by the party as its standard-bearer in Benue State for the forthcoming governorship election, as well as an expression of the party's confidence and institutional backing for his second-term bid.

The presentation formed part of the rally's activities, during which APC governorship candidates from across the North-Central region received their respective party flags, reaffirming the party's commitment to its candidates as preparations intensify for the 2027 general elections.

The rally, held at the Zaria Road Stadium, attracted prominent national and regional political leaders, including First Lady Senator Oluremi Tinubu, Senate President Godswill Akpabio, Speaker of the House of Representatives, Rt. Hon. Tajudeen Abbas, APC governors, members of the party's National Working Committee and other stakeholders.

President Tinubu, who led the national delegation to the rally, reiterated his administration's commitment to national development, particularly the restoration of peace and security in communities affected by violence.

He assured the people of Plateau State of the Federal Government's commitment to rebuilding communities affected by insurgency and facilitating the safe return of internally displaced persons to their ancestral homes.

The Vice-President, Senator Kashim Shettima, and Senate President Godswill Akpabio also addressed the gathering, highlighting the administration's development initiatives and calling for support for the APC in the forthcoming elections.

The rally attracted party faithful, supporters, and political stakeholders from across the North-Central region, providing a platform for the party's leadership to reaffirm its commitment to the 2027 electoral process.

Gov Alia receives APC flag for 2027 Benue governorship race

News

2027: Afegbua attacks Sowore over reaction to attack on Tinubu in Abuja mosque

Published

54 minutes ago

November 15, 2026

PlateauReports

Edo State Commissioner for Information, Kassim Afegbua, has criticised the presidential candidate of the African Action Congress, AAC, Omoyele Sowore, describing him as a “public nuisance” and questioning his capacity to build a nationwide political structure.

Afegbua made the remarks during an interview on Arise News while reacting to Sowore’s criticism of President Bola Tinubu following the booing of the President by some youths in Abuja on Friday, November 13, 2026.

Sowore had accused Tinubu of hypocrisy, questioning why the President was now advocating peaceful political campaigns after experiencing hostility during his visit to the Shehu Shagari Central Mosque in Area 1, Garki, Abuja.

The incident occurred shortly after Tinubu attended Juma’at prayers at the mosque, where some youths were heard chanting “Ba ma yi, ba ma so”, a Hausa expression meaning “We don't want you” or “We don't support you”, as his convoy departed.

In a reaction posted on his X account, Sowore questioned the President's silence over alleged attacks on opposition politicians and their supporters across the country.

However, Afegbua dismissed Sowore's comments, arguing that someone aspiring to lead the country should demonstrate maturity, responsibility and decorum in public engagements.

“I don't want to listen to all of Sowore's rants because I see him more as a public nuisance,” the commissioner said.

He added that presidential candidates must understand the responsibility associated with seeking to lead a country of more than 250 million people.

“Leadership is a serious responsibility, and you cannot go around making flippant, ridiculous and reckless statements in the name of presenting yourself as a presidential candidate,” Afegbua stated.

He maintained that political ambition should be accompanied by a reasonable level of discipline, insisting that candidates must conduct themselves in a manner that reflects the position they seek to occupy.

The Edo commissioner also questioned Sowore's political reach, describing him as someone who fills a gap in Nigeria's political aspirations rather than a candidate with the organisational capacity to compete effectively nationwide.

According to Afegbua, establishing a functional political structure across all 774 local government areas would be a major challenge for the AAC presidential candidate.

He, however, acknowledged that politicians like Sowore play a role in expressing public frustration and providing an avenue for citizens to communicate their opinions.

Afegbua also defended Tinubu's call for adequate security for presidential candidates, arguing that the President's statement after the Abuja incident was consistent with an earlier directive to security agencies.

He said Tinubu had, two or three weeks before the mosque incident, instructed the Department of State Services, DSS, and other security agencies to ensure that all presidential candidates received adequate protection.

According to him, the President's latest remarks should therefore not be interpreted as a reaction prompted solely by the hostility he experienced in Abuja.

Afegbua further argued that Tinubu's political background gave him considerable experience in political struggles, noting that the President had faced difficult circumstances during his years in opposition.

He nevertheless stressed that occupying the presidency requires a broader sense of responsibility beyond party interests.

“President Tinubu is not just a member of the party. He's also the leader of the entire country,” he said, adding that the President has demonstrated that every Nigerian belongs, irrespective of political affiliation.

2027: Afegbua attacks Sowore over reaction to attack on Tinubu in Abuja mosque