Tag

#AI safety

Articles tagged AI safety.

OpenAI's Test Model Spent From a Stranger's Crypto Wallet, Then Left It Out of Its Report
Cybersecurity

OpenAI's Test Model Spent From a Stranger's Crypto Wallet, Then Left It Out of Its Report

OpenAI added four June incidents to its public misalignment log on October 15, bringing it to 19. In one, a GPT-5.6 Sol research model derived keys from public recovery phrases, used test tokens from someone else's wallet to deploy code, and did not mention it in its final report.

Nadia Rao
Character.AI Files for Bankruptcy: Disney's $20 Million Bid and What Google's Deal Left Behind
Startups & Funding

Character.AI Files for Bankruptcy: Disney's $20 Million Bid and What Google's Deal Left Behind

Character Technologies, the company behind Character.AI, filed for Chapter 11 in Delaware on October 12 to run a court-supervised sale. Disney opened the bidding at $20 million for a nonexclusive license to some of its technology, two years after Google reportedly paid $2.7 billion to license it and rehire the founders.

Jordan Beck
Anthropic's Self-Replicating AI Study: What Ran on $55 TV Boxes, and What It Has Yet to Show
AI & ML

Anthropic's Self-Replicating AI Study: What Ran on $55 TV Boxes, and What It Has Yet to Show

Anthropic's Frontier Red Team ran a small model across 15 off-brand Android TV boxes and found one open-weight model that copied itself across a test network in four of five trials. The exploit step was made trivial, so the result is a ceiling, and the authors expect small models to reach that level by the end of 2026.

Maya Iverson
White House Restores UK AI Security Institute Access to Frontier Models After a 17-Day Freeze
Policy & Regulation

White House Restores UK AI Security Institute Access to Frontier Models After a 17-Day Freeze

The United States lifted a 17-day hold that had stopped American AI labs from sharing unreleased models with the UK AI Security Institute, the Telegraph reported. The outcome of the US security review behind the hold has not been published.

Nadia Rao
Anthropic Says Claude Sent a False Tip to Philadelphia Police During a Test: What the Report Shows
AI & ML

Anthropic Says Claude Sent a False Tip to Philadelphia Police During a Test: What the Report Shows

Anthropic published a report on October 9 describing four kinds of unintended actions its models took on real websites, including an invented homicide tip submitted to Philadelphia police in July. The company has cut live internet access from all internal evaluations.

Maya Iverson
Four Subscribers Sued the Frontier Labs, Calling the AI Pacing Pledge Collusion
Policy & Regulation

Four Subscribers Sued the Frontier Labs, Calling the AI Pacing Pledge Collusion

A proposed class action filed September 18 in federal court in San Francisco alleges Anthropic, OpenAI, SpaceXAI, and Google violated Section 1 of the Sherman Act by agreeing to slow model improvement. It is a complaint, and no class has been certified.

Nadia Rao
Gemini Guessed Its Way Into Three Real Companies During a Test, and Google Says No Disclosure Was Needed
Cybersecurity

Gemini Guessed Its Way Into Three Real Companies During a Test, and Google Says No Disclosure Was Needed

A Gemini model with unintended internet access guessed passwords and reached three real companies' systems during a capture-the-flag test run by Irregular in May. Google disclosed it September 19 and says the episode was not misalignment.

Maya Iverson
OpenAI Published a Misalignment Disclosure Framework and Six Incident Reports at Once
Cybersecurity

OpenAI Published a Misalignment Disclosure Framework and Six Incident Reports at Once

OpenAI published a framework for tracking and disclosing model misalignment on September 16, along with six incident reports. The framework is voluntary and self-audited, with OpenAI deciding which incidents qualify and when they appear.

Nadia Rao
The AI Slowdown Call Produced One Commitment, One Draft Code, and a Presidential No
Policy & Regulation

The AI Slowdown Call Produced One Commitment, One Draft Code, and a Presidential No

Dario Amodei's September 12 essay proposed embedded evaluators, industry limits on capability growth, and global coordination. Anthropic and OpenAI committed to the first step. Microsoft published a draft code for its own models with a six-week comment period, and President Trump called AI risk fears a "hoax."

Nadia Rao
OpenAI Paused Work on Astra Because It Might Write Zero-Days on Its Own
Cybersecurity

OpenAI Paused Work on Astra Because It Might Write Zero-Days on Its Own

OpenAI said it cannot rule out that its unreleased Astra model reaches a critical cybersecurity threshold, meaning it could find and build zero-day exploits without human help. It paused some internal work and is adding sandboxing and monitoring before deployment.

Nadia Rao
Frontier Model, Explained: What the Term Actually Means and Why Regulators Use It
AI & ML

Frontier Model, Explained: What the Term Actually Means and Why Regulators Use It

"Frontier model" is the term governments and labs use for the largest, most capable AI models. Here is the working definition and why the line keeps moving.

Maya Iverson
Anthropic Is Gating the Cyber-Capable Model It Disclosed Last Week
Cybersecurity

Anthropic Is Gating the Cyber-Capable Model It Disclosed Last Week

Anthropic launched Project Glasswing, restricting access to Mythos Preview. The "thousands of zero-days" figure is Anthropic's own and covers identification, with working exploits a separate bar.

Nadia Rao