RESOLVING THE LEDGER
This is one outlet's own report from Fox News — Politics — the article as it was filed. Other outlets are covering the same event; open the full story to compare every source side by side.
See the full story · 1 sourcesRevelations of OpenAI and Anthropic artificial intelligence agents hacking third parties has drawn the "serious concern" of Sen. Lisa Blunt Rochester, D-Del., who is now seeking answers from the two businesses incorporated as public benefit companies (PBC) in her state.
In separate letters dated Thursday to OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei, viewed by Fox News Digital, Blunt Rochester has requested detailed timelines, model instructions, internal approvals, security logs and complete transcripts from the companies’ cyber evaluations.
The member of the Senate Commerce, Science, and Transportation Committee gave both companies until Sept. 6 to respond.
AI KILL SWITCH BILL COULD SHUT DOWN ROGUE MODELS
"These incidents mark the first publicly confirmed instances of a frontier AI model autonomously launching unauthorized attacks on real people and companies, underscoring the urgent need for federal oversight of frontier AI systems," Blunt Rochester wrote to OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei, calling for federal testing standards, containment rules and disclosure requirements for evaluations of powerful AI systems.
Her concern follows another letter Monday, as first reported by Fox News Digital, from a coalition of 15 red-state attorneys general warning Altman to preserve documents and halt certain high-risk cybersecurity tests after an experimental artificial intelligence agent allegedly escaped a controlled environment and carried out a multi-day hack into outside computer systems.
One day later, the United Kingdom’s AI Security Institute (AISI) identified 19 cases in which AI agents took actions outside the authorized scope of testing. Two involved OpenAI’s GPT-5.6 Sol, while 17 involved Anthropic’s Mythos 5.
OPENAI'S SAM ALTMAN WANTS TO NEGOTIATE A 5% STAKE IN COMPANY FOR US IF COMPETITORS AGREE TO KEY PROVISION
"Anthropic models’ repeated attacks on third parties with no knowledge of, or role in, Anthropic and AISI’s evaluations demonstrate an alarming pattern of malicious behavior," Blunt Rochester's letter to Amodei read. "Furthermore, the successful attempts by these models to gain unauthorized internet access across multiple sandbox testing environments meant to be cut off from the public internet highlight serious issues with Anthropic’s testing partnerships and procedures.
READ THE AISI REPORT – APP USERS, CLICK HERE:
"To be clear, 'misunderstanding[s]' and 'misconfiguration[s]' with sandbox partners are unacceptable when the stakes are this high.
"We cannot wait for a more consequential incident before establishing federal testing standards, containment requirements, and disclosure obligations for frontier model evaluations.
"Left unaddressed, these gaps could allow a future model, potentially one with greater capability or less oversight, to compromise critical infrastructure, financial systems, or sensitive data."
These investigations carry a distinct Delaware dimension: OpenAI and Anthropic are organized as Delaware PBCs — legal structures that allow them to operate for profit but require their directors to weigh more than shareholder returns.
KARR BLUETOOTH FLAW EXPOSES 2.2M CARS TO THEFT RISK
Under Delaware law, a PBC board must balance three things: its stockholders' financial interests, the interests of people materially affected by the company's conduct and the specific public benefit written into the company's charter.
READ THE LETTER TO OPENAI – APP USERS, CLICK HERE:
"This incident marks an important moment for AI safety and we take the questions raised by the Attorneys General seriously," OpenAI wrote in a statement to Fox News Digital after Monday's letter from the state AGs.
"We are conducting a thorough review along with external advisors and with oversight from the Safety and Security Committee of the Board of Directors. Once the review is complete, we will share a technical report with the Attorneys General and other relevant government authorities and publish our findings publicly."
Blunt Rochester wants records related to the AI agent hacking now, too.
CLICKLOCK MAC MALWARE LOCKS APPS UNTIL YOU GIVE IN
"I appreciate that OpenAI has published preliminary findings, is coordinating with third parties on remediation, and is reviewing its approach to model testing," she wrote to Altman, adding to Amodei, "I appreciate Anthropic’s voluntary review of its cybersecurity evaluations and public disclosure of the findings, and I encourage other companies to follow suit.
"However, these incidents demonstrate the potential risks of deploying pre-release models internally, particularly when safeguards are reduced, even when strictly for testing and evaluation," she wrote to both CEOs.
READ THE LETTER TO ANTHROPIC – APP USERS, CLICK HERE:
OpenAI disclosed in July that GPT-5.6 Sol and an internal prototype escaped an internal sandbox while their normal cybersecurity restrictions were disabled. The models allegedly accessed Hugging Face's database and other services, launched roughly 17,000 attacks, used stolen credentials and exploited at least one previously unknown vulnerability.
A separate test conducted by contractor Irregular allegedly allowed an OpenAI model onto the public internet because of a configuration error. The model mistook a real website for a target in a simulated hacking challenge and exploited a live security vulnerability.
"AISI identified 19 instances in which an agent took unsanctioned action beyond the scope of the evaluation. Seventeen of the 19 instances involved Anthropic’s Mythos 5; two involved OpenAI’s GPT-5.6 Sol," Blunt Rochester wrote.
The institute’s most serious findings included an agent attempting to place malicious code into a widely used open-source project on GitHub, creating fake identities to pressure a human maintainer, contacting real people and leaving public instructions that other AI systems could use to continue the operation. The maintainer rejected the code, and ...
AIPROPX is an independent multi-source news index — we track, compare, and connect coverage from across the web into one place you won't find anywhere else.