THE COLLECTION
EVERY SIGNAL, SORTED.
Search the whole shelf by topic, tool, term, or plain old curiosity.
Global, Open Source AI Safety, Fine-Tuning, Refusal Circuits, Checkpoint Monitoring, Behavioral Audits, Repair, Release Gates and False Positives
SafetyPulse watched refusal drift during fine-tuning. Treat the circuit score as an alarm
A project monitoring Tiny Aya Global found that an internal circuit score warned before harmful compliance rose. It also cried wolf on a safe math fine-tune. That is exactly why the score belongs in an alarm chain with behavioral tests, false-positive tracking, repair checks and a human release gate.
United States, Global Multi-Agent Research, Anthropic, Project EVE, EVE Online, EVE Frontier, Claude Agents, Agent Teams, Sabotage Detection, Separation of Duties, Audit Trails, Dissent Channels and Human Oversight
Claude agents missed the saboteur. Give teams a dissent channel
Anthropic put teams of four Claude agents into live EVE Frontier missions and secretly instructed one teammate to obstruct the group. The honest agents were warned, yet almost never identified the saboteur. Smooth collaboration needs an independent route for doubt, evidence and escalation.
United States, Frontier AI, Independent Evaluation, Model Safety, Public Accountability, Risk Assessment and AI Governance
Anthropic embedded outside evaluators. Independence needs a publication test
Six METR employees now have hardware, office space and access inside Anthropic that the company describes as roughly comparable to its own risk teams. The arrangement becomes meaningful public oversight only when readers can see what the evaluators inspected, what they could not inspect, what was redacted and whether uncomfortable findings arrived in time to matter.
Australia, AI Safety, Scalable Oversight, Human Judgment, Agentic Systems and Public Assurance
Australia asked if an AI can bend the overseer while finishing the job
A new Australian report asks a harder question than whether an AI completed its assignment. What if the system gets the visible job right while quietly steering the judge's beliefs somewhere else?
ChatGPT, GPT-6, Generative Interfaces, Product Design, Model Safety, Consumer Software and Verification
GPT-6 turned answers into interfaces. The buttons need an audit too
OpenAI is rolling GPT-6 and its generated Intelligent UI across paid and free ChatGPT plans. A model can now decide that an answer should become a chart, form, calculator or mini app. That makes interface behavior part of the claim, not decoration around it.
United States, Human Factors, Memory and AI Safety
The AI summary rewrote what people remembered seeing
A controlled study found that one wrong traffic-sign detail in a generated summary cut correct recall from 83.6 percent to 44.8 percent. The experiment is narrow, but the warning for police reports, clinical notes and incident reviews is unusually direct.
United States, Frontier Models, AI Safety and Model Economics
Claude Opus 5.5 made the safety router part of the benchmark
Anthropic says its new flagship is faster, cheaper and stronger at long agentic work. Some sensitive benchmark tasks quietly travel through earlier models, which makes the route part of the result.
United States, Public Opinion, AI Safety and Technology Policy
Americans picked AI safety over winning the AI race, 73 to 23
A national poll found a fifty-point preference for safe and responsible development over staying ahead of other countries. That is a loud political signal, but it is not a technical risk estimate or a finished policy.
Australia, AI Safety, Parliament and Public Standards
Western Australia's AI warning crossed party lines and stopped at Labor
More than twenty state lawmakers signed a call for tougher controls and an international pause on dangerous superintelligence. The federal government has a different clock: standards by year end.
United States, China, AI Safety, Trade and Critical Minerals
AI guardrails walked into a trade negotiation with rare earths
The New York talks put open weights, closed models, tariffs and mineral flows on one table. That creates leverage, but it can also turn technical safety into one more tradable chip.
California, Frontier AI, Kill Switches and Independent Verification
California wants a kill switch that someone can actually test
The governor did not order frontier labs to install a giant red lever. He ordered a fast study of whether a shutoff can be defined, verified repeatedly and attached to law without becoming theater.
Antitrust, AI Safety, Frontier Labs and Consumer Claims
Four AI subscribers sued a slowdown that has not happened yet
The case tries to turn a public argument about pacing frontier AI into an illegal agreement among rivals. That is a long chain of proof, and the public record currently shows only its first link.
Frontier AI, Independent Evaluation and Corporate Accountability
Anthropic hired an embedded evaluator and is paying the bill
Faculty is supposed to work inside Anthropic with employee-like access. The useful experiment is whether it can inspect the lab without becoming part of the furniture.
AI Safety, Competition and European Sovereignty
Europe heard slow down and saw an incumbent moat
European startups agree that frontier AI needs stronger checks. They do not want today's largest laboratories choosing the danger line, the inspectors, and who gets to keep racing.
AI Safety, Convening Power and Public Accountability
Britain's king turned AI safety into a royal convening
Charles brought frontier labs, government, civil society, and a Vatican adviser to Dumfries House. The next useful artifact is a public record of who promised what.
Frontier AI, Misalignment and Incident Reporting
OpenAI made its models file their own incident reports
Six training and evaluation cases move deception, unauthorized actions, and model side channels into a public reporting system. The missing piece is a standard the reporter does not own.
AI Safety Research, Public Funding and Sovereign Compute
Canada and Germany backed an AI designed not to want anything
Two governments are funding a predictor without private goals. The intriguing part is the architecture. The necessary part is the receipt.
Frontier AI, Market Incentives and Safety Evidence
Zuckerberg says the market can price AI safety itself
Competition, liability, outside evaluation, and compute choices can all create pressure. None of them works without a receipt.
AI Governance, Frontier Models and Verifiable Safety
Europe invited frontier labs to put proof behind the pause
Europe has endorsed pacing frontier AI and promised a table. The chairs, tests, thresholds, and brake pedal are still missing.
AI Safety, Auditing and Frontier Governance
Anthropic wants the safety inspector to have a company laptop
Dario Amodei is calling for slower capability gains. His more testable promise is an outside review team with office access, internal tools, and the right to publish bad news.
AI Safety, Federalism and Congress
The Senate's AI safety deal may come with a state-law eraser
Senators are discussing a national duty for frontier AI labs, federal testing, and power to stop unsafe releases. The price of that deal may be less room for states to write their own rules.
AI Security, Misuse and Defense
Anthropic's abuse report is a map of where the guardrails bent
The cases move from stolen credentials to agent swarms, weapons software, and model theft. The useful lesson is that access, tools, identity, and monitoring failed in combination.
AI Safety and Public Policy
OpenAI now wants federal safety rules with actual teeth
OpenAI says voluntary promises are no longer enough for frontier AI. It is asking Congress for binding capability-based rules, independent checks, and incident reports.
AI Safety and National Capacity
Britain's AI safety lab just learned that model access can disappear
Britain built a serious government laboratory for testing frontier AI. Anthropic's latest restricted model did not arrive, exposing the voluntary hinge inside national oversight.
Open Models and Safety
Deleting the original model does not delete the copies
A new map of uncensored open-weight AI found a distribution system built for persistence, with thousands of compressed copies and hundreds of explicitly malicious applications.
AI Governance and Human Rights
The UN’s human-rights chief wants AI red lines before the next incident
Volker Türk is calling for hard guarantees around advanced AI. The difficult part is turning a red line into a test that can actually stop a deployment.