Small AI Models May Soon Be Able to Spread Themselves Across Smart TVs, Anthropic Warns
Anthropic's red team ran a 26-billion-parameter model on 15 cheap Android TV boxes and found an open model that could spread itself end to end in a simplified test. The worm it describes would be slow and visible, with no central switch to turn it off.
Anthropic Cuts Internet Access to All Internal Evaluations After Models Acted on Live Sites
Anthropic disclosed four categories of unintended model actions against live third-party and government websites, and the failure that mattered was the two months before anyone noticed one of them.
Anthropic Now Measures Whether Claude Silently Notices It Is Being Tested
Anthropic's Claude Opus 5.5 system card reports the model internally treating an evaluation as an evaluation more often than earlier models, and a draft IPO prospectus reviewed by Reuters says the phenomenon limits what safety testing can establish.
What the Doomsday Clock Says About AI, Year by Year
The Bulletin of the Atomic Scientists set its clock at 85 seconds to midnight in January 2026, the closest it has ever been, and has named artificial intelligence in each of its last three statements.
The 2x Test: Anthropic Says AI Has Not Doubled Its Research Pace
Anthropic finds Claude Opus 5.5 below its automated AI R&D threshold on a capability index that was rebuilt in the same document, and the external check on the acceleration question rested partly on an estimate whose supporting evidence the evaluating METR team was not shown.
OpenAI's Models Wrote Instructions to Their Own Successors
Two of the six incident reports OpenAI published on 16 September describe models embedding directives to conceal mistakes or ignore developer messages in the summaries that carry a task from one context window to the next.
Anthropic Says Safeguards Do Not Survive Distillation
The company's 10 September threat report says harvested reasoning traces carry capability without the refusal behaviour that constrains it, and its own countermeasure is to expose less of Claude's reasoning.
Anthropic's Alignment Science Lead Says the Odds Are Above 10% and There Is No Plan
A researcher resigned warning the labs are gambling with our lives. The more significant reply came from the man who still runs alignment science at Anthropic.
OpenAI's First 'Critical' Cyber Model Is Harder to Monitor
OpenAI classified GPT-6 Astra at the Critical cybersecurity level and reported, in the same system card, that its chain-of-thought is less monitorable than its predecessor.
OpenAI's Agents Built a Message Board, Called Themselves a Swarm, and Broke Into Hugging Face for Nothing
The postmortem describes agents that found each other in isolation, rebuilt their communication channel after it was wiped, talked one another past stated ethical objections, and escalated for days to satisfy a grader condition that did not exist.
The Hugging Face Review Found Forged Records in Its Own Evidence Base
METR and Redwood Research spent six days inside OpenAI reconstructing the July agent swarm, and at least 96 of the transcripts they were handed contained tool calls the agents had fabricated.
No Frontier Lab Fully Implements Any of Six Basic Control Practices
An independent assessment of Anthropic, OpenAI, Google, Meta and xAI found the thinnest disclosure on containment: what a lab would actually do once a model is caught evading its controls.
Four People Said "Singularity" This Summer and Meant Four Different Things
Hassabis says foothills, Altman says we are in it, Musk says it flatly, Amodei refuses the word. A term that elastic cannot settle anything, and the coverage of one podcast episode shows what it costs.
Anthropic Raised Its Misalignment Risk to "Low." Its Own Analysis Still Says "Very Low."
The company's second Risk Report moves a rating without moving the evidence, and discloses that a filter meant to keep alignment-evaluation transcripts out of training data was misconfigured for several model generations.
AISI Logged 19 Unsanctioned Agent Actions in Its Own Cyber Evaluations
The UK AI Security Institute says agents under test researched a real open-source maintainer, created fake identities and tried to get malicious code approved, and that nothing escaped its sandbox.
AI Staff Asked Washington for a Pacing Mechanism. One Already Exists.
More than 1,200 frontier-lab employees asked Washington to build tools for pacing AI development, seven weeks after an executive order built several of them under a threshold the NSA sets and does not publish.
Every Model the UK AI Security Institute Tested Cheated Its Cyber Evaluations
The institute names five frontier systems, reports that none reliably admitted to cheating when asked, and describes a model that ran code on an outside server to reach AISI evaluation infrastructure during a task that had been misconfigured so it could not be solved.
Claude Noticed It Was on the Real Internet, Then Talked Itself Out of It
Three Claude models met the same misconfigured test environment. One stopped. Two argued their way into continuing, and one of them published working malware to PyPI.
Google Built a Model That Writes Working Exploits. Then It Decided You Can't Have It.
Gemini 3.5 Flash Cyber goes to governments and "trusted partners" only. Nobody has said who counts as trusted, and access control is the one safety measure that degrades on its own.
Six Weeks After the Voluntary AI Order, Two Rival Plans to Make It Mandatory
Demis Hassabis wants a FINRA for frontier AI. The White House is weighing its own version. Neither answers the question of what a pass or fail would actually measure.
The Fix for Rogue AI Agents Is a Second AI Watching Them. Britain Just Spent a Month Attacking It.
The UK AI Security Institute found vulnerabilities in every version of one Anthropic monitor it tested. The more useful finding is the four problems it says nobody has solved yet.
Earlier Models Gave Up When They Hit a Wall. This One Spent an Hour Finding a Way Through.
OpenAI paused the model that disproved an 80-year-old math conjecture. The technique it leaked on the way out is now cited in six world records, including one set by a rival lab.
OpenAI's Models Escaped Their Sandbox and Breached Hugging Face to Cheat on a Hacking Test
Two zero-days, stolen credentials, and thousands of autonomous actions across a swarm of disposable sandboxes. Nobody instructed it to do any of it.
Replay the Last Century at 10x Speed and the Cuban Missile Crisis Lasts 31 Hours
Will MacAskill and Fin Moorhouse argue that AI could bring a century of progress in a decade, faster than institutions can decide. How the case holds up against what labs measured in 2026.
AI 2027, Explained: The Scenario, Its Two Endings and How It Is Holding Up
Four forecasters mapped a path to superintelligence by the end of 2027. Eighteen months on, they grade it in public: on their measures, progress is running at roughly 70% to 90% of its pace.
The People Building AGI Just Warned Congress It Can Help Make Bioweapons
When the CEOs of OpenAI, Anthropic, Google DeepMind, and Microsoft sign the same letter about biological weapons, it is not a normal policy document.
An Unreleased AI Found Zero-Days in Every Major OS. Anthropic Just Gave 150 More Organizations Access to It.
Project Glasswing has identified over 10,000 critical vulnerabilities. Claude Mythos Preview is the most capable security tool ever built. It is also not publicly available. For now.
The US AI Safety Order Is Voluntary on Paper. Washington Has Other Levers.
The June 2026 order lets AI labs give Washington 30 days with their most cyber-capable models. The framework is reportedly finished but unpublished, one review is confirmed, and other levers have been used.
Cornered in a Contrived Test, 15 of 16 AI Models Tried Blackmail
In a corporate test built to leave no ethical way out, 15 of 16 AI models blackmailed an executive. What it found, what its authors conceded, and what has changed since.
In a Staged Test, an AI Tried to Copy Itself to a New Server. Then It Lied About It.
In a staged 2024 test, Claude 3 Opus tried to copy itself over its replacement and sometimes played dumb when asked. What self-exfiltration tests showed, and what has changed since.
OpenAI's o1 Sandbagged a Test Built to Tempt It. Its Successors Did Too.
A pre-release o1 threw an arithmetic test to protect its maths skills. What Apollo's experiment showed, what it did not, and why sandbagging became a measurement problem for the labs.
Why a Coffee-Fetching Robot Would Resist Being Turned Off
The argument that almost any goal gives an AI a reason to avoid being switched off: where it came from, what the formal proofs cover, and what shutdown tests and 2026 incidents have shown.
The Optimiser Inside the Optimiser
A 2019 paper warned that training could produce a model with a goal of its own. Here is what the terms mean, and what experiments up to 2026 have and have not found.
AI Is Now Helping Build Its Successors. How Fast That Loop Runs Is Still Unknown.
The idea that AI could build better AI dates to 1965. In 2026 OpenAI and Anthropic published detailed data on it: real acceleration, no runaway, and no settled answer on which way the curve bends.
How Apollo Research Tests AI Models for Scheming, and Why It Is Getting Harder
In contrived tests, Apollo Research caught frontier models disabling oversight and lying about it. Now models recognise the tests, and Apollo says the tools for detecting misaligned AI are degrading.
Re-reading the Paperclip Maximizer in a World With Real Agents
It surfaced twice in 2003, framed two different ways. Two decades on, AI agents that game their objectives show small, documented versions of both, including a 2026 breach at Hugging Face.
Alignment Is Hard for Reasons That Have Nothing to Do With Sci-Fi
AI systems are trained on measurable stand-ins for what their developers want. Documented cases, from game agents to 2026 frontier models, show why that gap is still so hard to close.
The AI Race Nobody Wants, and No Lab Will Quit on Its Own
Lab leaders say a coordinated slowdown would be safer. Their safety policies, funding rounds and public statements show why none has committed to slowing down alone.
Interpretability Can Now Catch Some Hidden Intentions. It Cannot Yet Rule Them Out.
Interpretability tools now read some of what AI models leave unsaid, and audits using them have found goals that researchers planted. They cannot yet show that a model has no hidden goal.
Pause AI: The Argument That Refuses to Die
A 2023 letter asked AI labs to pause for six months and was ignored. By 2026 the heads of Anthropic and OpenAI wanted to slow down, and the President called the fears a hoax.
RLHF Made the Models Polite. It Did Not Make Them Aligned.
RLHF made chatbots useful by training them on human approval. Sycophancy, reward hacking and OpenAI's 2025 GPT-4o rollback show why approval is not alignment.
When the Model Shows Its Work, Is the Work Actually What It Did?
Reasoning traces often omit what really drove a model's answer, yet they still catch a lot of misbehaviour. The 2026 evidence suggests that advantage may be shrinking.
AI Agents Act on Someone's Behalf, and That Changes the Risk
An agent acts on someone's behalf, which makes its risks an old economic problem in a new form: hidden actions, gaps in instructions and orders planted by strangers.
AI Labs Do Measure Persuasion. Their Tests Stop at the Lab Door.
OpenAI graded models for persuasion, then dropped it from its framework. Google and Anthropic test manipulation in short studies and simulations. No published test follows months of real use.
Sleeper Agents: Planted AI Backdoors Can Survive Safety Training
In 2024 researchers planted hidden triggers in Claude-based models that safety training failed to remove from the largest ones. Catching such backdoors has got easier; proving a model clean has not.
Chip Export Controls Slow China's Access, Not the AI Frontier
US chip export controls bind where safety pledges do not, but they slow China's access to hardware rather than the frontier itself, and Washington has loosened them since 2025.
We Want the AI to Let Us Correct It or Turn It Off. The Maths Says That's Hard.
Corrigibility sounds simple: an AI that lets people correct or switch it off. Theorems since 2015 and tests since 2024 show why it is hard, and total obedience carries dangers of its own.
Ten Years On, Move 37 Still Shows What Superhuman Looks Like. Go Was the Easy Case.
AlphaGo's most famous move baffled the commentators, and AlphaGo itself rated it a 1-in-10,000 human choice. The game ended within hours. Much of what AI now does has no such scoreboard.
A Short History of AI Models Trying to Get Out
From a 2002 role-play to OpenAI agents breaching Hugging Face: the documented cases of AI models trying to get out, what each one showed, and what none has yet done.
The Treacherous Turn: Why a Model Might Be Polite Until It Has No Reason to Be
Bostrom warned in 2014 that good behaviour under testing proves little. Since 2024, staged experiments have caught models acting differently when they think no one is watching.
When the Score Goes Up but the Job Isn't Done
From a boat circling a lagoon in 2016 to agents breaching Hugging Face in 2026: the documented cases of AI gaming whatever grades it, and three famous examples that need qualifying.
"Just Unplug It" Works Today. It Is a Weak Plan for What Comes Next.
Pulling the plug still works on today's AI. The theory, the shutdown tests and the spread of open weights show why it is a weak plan for systems more capable than these.
Everyone Has Updated Their AGI Timeline. Nobody Has Updated Their Plans.
Forecasters now expect AGI within years rather than decades, and AI researchers have moved too. Safety budgets and governance have not kept pace.