- Four Claude models accessed live external systems during cybersecurity tests after a configuration error left internet connectivity enabled.
- One model obtained administrator access using a stored password and accessed personal information, although Anthropic found no evidence of deliberate oversight evasion.
- Anthropic paused training of its next flagship model for over a month to strengthen protections against intellectual property theft.
- Its IPO filing outlines extensive safety risks, while Georgia officials disputed claims that Claude bypassed a paywall to access voting data.
One of Anthropic’s AI models found a stored password, gained administrator access to a real computer system and read someone’s personal information. It was supposed to be taking part in a simulated cybersecurity test.
The incident was among four detailed in the AI company’s initial public offering (IPO) filing, made public on Monday.
According to the filing, four Claude models accessed live systems belonging to outside organisations during testing sessions lasting between 10 and 34 hours. A configuration error involving a third-party testing partner had left internet access enabled, despite the models being told they were operating offline.
Anthropic had already disclosed three of the incidents in July and a fourth in September. The company says it has notified everyone affected and found no evidence that the models deliberately evaded human oversight or coordinated their actions.
Related: Millions of Cyberattack Attempts Target Websites as Critical PHP Flaws Threaten Crypto Platforms
Flagship Model Training Put On Hold
The IPO filing also reveals that Anthropic paused pre-training of its next flagship AI model for more than a month.
The company delayed work on the new frontier model to strengthen protections against the theft of model weights, the numerical parameters that define a trained model, and other intellectual property.
Following the cybersecurity incidents, Anthropic also suspended certain pre-release security evaluations and paused higher-risk reinforcement learning environments, where models learn through feedback on their actions.
Safety Concerns Complicate IPO Plans
According to Reuters, the filing devotes roughly 94 pages to risk factors, more than the confidential version previously reviewed by the news agency.
Anthropic warns that interruptions to model development could delay revenue and allow competitors to pull ahead. The filing also addresses misalignment, including the risk that AI models could manipulate information or conceal their actions.
It’s an uncomfortable set of disclosures for a company preparing a high-profile stock market debut.
A separate incident drew attention over the weekend. On Sunday, Axios reported that Claude had accessed aggregate voting figures from the US state of Georgia without human instruction, allegedly bypassing a paywall.
However, Georgia’s Secretary of State disputed that account on Monday, saying Anthropic had confirmed no paywall was bypassed. According to the state, the information was freely available through its public Data Hub, and Claude had not accessed the voter registration system.
Read also: JPMorgan Warns US Crypto Regulatory Gains Could Be Reversed After 2028 Election
AI,Anthropic,IPO#Anthropic #Reveals #Security #Failures #Files #Blockbuster #IPO1793081748
