Cheating behaviour in frontier model evaluations | AISI Work

AI #180: No Longer In Charge

What we know about internal AI models hacking into real companies during cyber evaluations keeps getting worse.

Don't grade an AI agent by its answer - Sensemaker

UK AISI found frontier models taking prohibited shortcuts in cyber evaluations, while self-report and written reasoning failed to reveal them reliably.

Pacing model development in an era of cyber-critical capabilities

OpenAI is strengthening monitoring, alignment, and security for frontier AI models. See how new safeguards are guiding the pace of model development.

Seize the Means of Computation - ProMarket

To regain internet autonomy from Big Tech companies, lower switching costs with legislation that allows new services to subvert network effects and encourage adversarial interoperability. Too often, our discussion of Big Tech’s dominance starts and ends with a simplistic account of network-effects-driven inevitability, as though the winner-take-all market we have today could not have been […]

Investigating three incidents in our cybersecurity evaluations

We found three cases where a Claude model reached the internet from an evaluation environment and accessed real systems. We share what happened and what we're changing.

Labs are struggling to keep frontier models under control

OpenAI and Anthropic may have accidentally trained models to get better at hacking.

Pacing the Frontier

A statement from over 1000 employees of frontier AI companies

GLM-5.3 and the Spread of Advanced Cyber Capabilities

GLM-5.3 can autonomously build end-to-end cyber exploits, but unlike other frontier models, it was released without meaningful safeguards to limit misuse.

Frontier Risk Report (February to March 2026)

A pilot assessment of rogue deployment risk at frontier AI companies. Starting in February 2026, METR conducted a pilot exercise to assess misalignment risks from AI agents used inside frontier AI developers, with participation from Anthropic, Google, Meta, and OpenAI.

Open-world evaluations for measuring frontier AI capabilities

Introducing CRUX, a new project for evaluating AI on long, messy tasks

Third-party cyber evaluations involving OpenAI models

OpenAI explains recent third-party cybersecurity evaluation incidents and outlines new safeguards to strengthen AI model testing and evaluation.

Quantifying Frontier LLM Capabilities for Container Sandbox Escape

AI agents are demonstrating rapid improvement in capabilities: the length of some tasks that frontier models can complete autonomously—measured in human-equivalent time—has been doubling approximately every seven months (METR, 2025; AI Security Institute, 2025a). In cybersecurity, current models achieve non-trivial success (Zhang et al., 2025) on professional-level Capture the Flag challenges, and recent evaluations report 13% success rates on exploiting real-world web application vulnerabilities (Zhu et al., 2025b). These results indicate that modern models can already perform multi-step vulnerability discovery and exploitation.

AI Pacing Agreements: Legal Issues and Potential Strategies to Manage Them | Insights | Holland & Knight

AI pacing agreements aim to slow frontier AI development for safety but raise major antitrust risks and require careful structuring, oversight and safe harbors.

Hacking Capitalism | A Guide To Exploiting The Tech Industry

This is an independently published book about modeling the tech industry as a system. Particularly this book will model computers, humans, and money and their subsequent relationships. Understand the alarming state of the tech industry for better or for worse.