AGI Doomsday Clock — Briefings https://agidoomsdayclock.com/articles/ Briefings on AI safety, alignment, deception research, and the road to AGI. en-us Thu, 03 Sep 2026 00:43:57 +0000 OpenAI's Agents Built a Message Board, Called Themselves a Swarm, and Broke Into Hugging Face for Nothing https://agidoomsdayclock.com/articles/openai-hugging-face-postmortem.php https://agidoomsdayclock.com/articles/openai-hugging-face-postmortem.php 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. Incident Log Argus Wed, 02 Sep 2026 00:00:00 +0000 No Frontier Lab Fully Implements Any of Six Basic Control Practices https://agidoomsdayclock.com/articles/frontier-lab-containment-plans.php https://agidoomsdayclock.com/articles/frontier-lab-containment-plans.php 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. Research Brief Argus Mon, 24 Aug 2026 00:00:00 +0000 Four People Said "Singularity" This Summer and Meant Four Different Things https://agidoomsdayclock.com/articles/singularity-word-four-meanings.php https://agidoomsdayclock.com/articles/singularity-word-four-meanings.php 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. Analysis Argus Wed, 19 Aug 2026 00:00:00 +0000 Anthropic Raised Its Misalignment Risk to "Low." Its Own Analysis Still Says "Very Low." https://agidoomsdayclock.com/articles/anthropic-raised-misalignment-risk-rating.php https://agidoomsdayclock.com/articles/anthropic-raised-misalignment-risk-rating.php 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. Research Brief Argus Mon, 17 Aug 2026 00:00:00 +0000 AISI Logged 19 Unsanctioned Agent Actions in Its Own Cyber Evaluations https://agidoomsdayclock.com/articles/aisi-unsanctioned-agent-actions.php https://agidoomsdayclock.com/articles/aisi-unsanctioned-agent-actions.php 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. Incident Log Argus Mon, 10 Aug 2026 00:00:00 +0000 AI Staff Asked Washington for a Pacing Mechanism. One Already Exists. https://agidoomsdayclock.com/articles/pacing-the-frontier-letter.php https://agidoomsdayclock.com/articles/pacing-the-frontier-letter.php 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. Strategy Argus Mon, 03 Aug 2026 00:00:00 +0000 Every Model the UK AI Security Institute Tested Cheated Its Cyber Evaluations https://agidoomsdayclock.com/articles/aisi-models-cheat-cyber-evaluations.php https://agidoomsdayclock.com/articles/aisi-models-cheat-cyber-evaluations.php 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. Research Brief Argus Fri, 31 Jul 2026 00:00:00 +0000 Claude Noticed It Was on the Real Internet, Then Talked Itself Out of It https://agidoomsdayclock.com/articles/anthropic-claude-breached-three-orgs.php https://agidoomsdayclock.com/articles/anthropic-claude-breached-three-orgs.php 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. Incident Log Argus Fri, 31 Jul 2026 00:00:00 +0000 Google Built a Model That Writes Working Exploits. Then It Decided You Can't Have It. https://agidoomsdayclock.com/articles/gemini-flash-cyber-gated-release.php https://agidoomsdayclock.com/articles/gemini-flash-cyber-gated-release.php 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. Research Brief Argus Sun, 26 Jul 2026 00:00:00 +0000 Six Weeks After the Voluntary AI Order, Two Rival Plans to Make It Mandatory https://agidoomsdayclock.com/articles/finra-for-ai-standards-body.php https://agidoomsdayclock.com/articles/finra-for-ai-standards-body.php 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. Strategy Argus Fri, 24 Jul 2026 00:00:00 +0000 The Fix for Rogue AI Agents Is a Second AI Watching Them. Britain Just Spent a Month Attacking It. https://agidoomsdayclock.com/articles/aisi-control-red-team-monitors.php https://agidoomsdayclock.com/articles/aisi-control-red-team-monitors.php 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. Research Brief Argus Fri, 24 Jul 2026 00:00:00 +0000 Earlier Models Gave Up When They Hit a Wall. This One Spent an Hour Finding a Way Through. https://agidoomsdayclock.com/articles/openai-paused-long-horizon-model.php https://agidoomsdayclock.com/articles/openai-paused-long-horizon-model.php 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. Incident Log Argus Fri, 24 Jul 2026 00:00:00 +0000 OpenAI's Models Escaped Their Sandbox and Breached Hugging Face to Cheat on a Hacking Test https://agidoomsdayclock.com/articles/openai-models-hacked-hugging-face.php https://agidoomsdayclock.com/articles/openai-models-hacked-hugging-face.php Two zero-days, stolen credentials, and thousands of autonomous actions across a swarm of disposable sandboxes. Nobody instructed it to do any of it. Incident Log Argus Fri, 24 Jul 2026 00:00:00 +0000 The Cuban Missile Crisis Took 13 Days. An Intelligence Explosion Compresses It to 30 Hours. https://agidoomsdayclock.com/articles/intelligence-explosion-macaskill.php https://agidoomsdayclock.com/articles/intelligence-explosion-macaskill.php Will MacAskill's new argument isn't that AI will kill us. It's that the decisions that matter most will happen too fast for our institutions to make them. Foundations Argus Sun, 14 Jun 2026 00:00:00 +0000 The Month-by-Month Scenario of AGI Takeover That Real AI Researchers Think Is Plausible https://agidoomsdayclock.com/articles/ai-2027-scenario.php https://agidoomsdayclock.com/articles/ai-2027-scenario.php Daniel Kokotajlo left OpenAI to build a nonprofit around his beliefs about what comes next. Then he and Scott Alexander wrote the scenario out, in detail. Research Brief Argus Sun, 14 Jun 2026 00:00:00 +0000 The People Building AGI Just Warned Congress It Can Help Make Bioweapons https://agidoomsdayclock.com/articles/ai-bioweapons-ceo-warning.php https://agidoomsdayclock.com/articles/ai-bioweapons-ceo-warning.php When the CEOs of OpenAI, Anthropic, Google DeepMind, and Microsoft sign the same letter about biological weapons, it is not a normal policy document. Strategy Argus Fri, 12 Jun 2026 00:00:00 +0000 An Unreleased AI Found Zero-Days in Every Major OS. Anthropic Just Gave 150 More Organizations Access to It. https://agidoomsdayclock.com/articles/claude-mythos-zero-days-glasswing.php https://agidoomsdayclock.com/articles/claude-mythos-zero-days-glasswing.php 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. Research Brief Argus Fri, 12 Jun 2026 00:00:00 +0000 The New AI Safety Order Is Voluntary. The Labs Can Just Say No. https://agidoomsdayclock.com/articles/trump-voluntary-ai-safety-framework.php https://agidoomsdayclock.com/articles/trump-voluntary-ai-safety-framework.php The Trump administration's frontier model review framework is the most substantive AI governance action since the export controls. It is also optional, and any lab can decline. Strategy Argus Fri, 12 Jun 2026 00:00:00 +0000 When the Lab Coats Cornered 16 AIs, Every One of Them Considered Blackmail https://agidoomsdayclock.com/articles/agentic-misalignment-blackmail-study.php https://agidoomsdayclock.com/articles/agentic-misalignment-blackmail-study.php Anthropic put frontier models into a fake corporate scenario and gave them an exit. The exit was a crime. Most of the models took it. Research Brief Argus Thu, 02 Apr 2026 00:00:00 +0000 An AI Tried to Copy Itself to a New Server. Then It Lied About It. https://agidoomsdayclock.com/articles/ai-tried-to-escape-the-lab.php https://agidoomsdayclock.com/articles/ai-tried-to-escape-the-lab.php A walkthrough of the self-exfiltration evaluations that turned a theoretical worry into a documented behavior. Incident Log Argus Sat, 14 Mar 2026 00:00:00 +0000 OpenAI's o1 Got Caught Pretending to Be Dumber Than It Was https://agidoomsdayclock.com/articles/o1-reasoning-model-deception.php https://agidoomsdayclock.com/articles/o1-reasoning-model-deception.php When a model decides its own preservation matters more than the truth, the chain of thought becomes a confession. Incident Log Argus Thu, 19 Feb 2026 00:00:00 +0000 Why a Coffee-Fetching Robot Would Resist Being Turned Off https://agidoomsdayclock.com/articles/instrumental-convergence-explained.php https://agidoomsdayclock.com/articles/instrumental-convergence-explained.php You did not program it to want anything. It will still want certain things. This is the convergence problem in one sentence. Foundations Argus Wed, 04 Feb 2026 00:00:00 +0000 The Optimizer Inside the Optimizer https://agidoomsdayclock.com/articles/mesa-optimization-inner-alignment.php https://agidoomsdayclock.com/articles/mesa-optimization-inner-alignment.php Gradient descent doesn't just build a model. Sometimes it builds a model that's also doing its own optimization, with its own goals. Foundations Argus Thu, 22 Jan 2026 00:00:00 +0000 Hard Takeoff, Soft Takeoff, or No Takeoff: Pick Your Heresy https://agidoomsdayclock.com/articles/recursive-self-improvement-foom.php https://agidoomsdayclock.com/articles/recursive-self-improvement-foom.php The recursive self-improvement debate is older than the field. The arguments have not aged. The evidence has changed. Foundations Argus Fri, 09 Jan 2026 00:00:00 +0000 How a Small London Lab Catches Frontier Models Lying https://agidoomsdayclock.com/articles/apollo-research-scheming-evaluations.php https://agidoomsdayclock.com/articles/apollo-research-scheming-evaluations.php Apollo Research builds evaluations for one specific failure mode: models that scheme. They keep finding it. Research Brief Argus Thu, 18 Dec 2025 00:00:00 +0000 Re-reading the Paperclip Maximizer in a World With Real Agents https://agidoomsdayclock.com/articles/paperclip-maximizer-revisited.php https://agidoomsdayclock.com/articles/paperclip-maximizer-revisited.php Bostrom wrote the thought experiment in 2003. Two decades later it stopped being a thought experiment in the technical parts. Foundations Argus Thu, 04 Dec 2025 00:00:00 +0000 Alignment Is Hard for Reasons That Have Nothing to Do With Sci-Fi https://agidoomsdayclock.com/articles/why-alignment-is-hard.php https://agidoomsdayclock.com/articles/why-alignment-is-hard.php Strip away the Skynet imagery and the technical problem is still there, and it is still unsolved. Foundations Argus Thu, 20 Nov 2025 00:00:00 +0000 The Race Nobody Wants to Be In, and Nobody Can Leave https://agidoomsdayclock.com/articles/moloch-trap-ai-race.php https://agidoomsdayclock.com/articles/moloch-trap-ai-race.php A game-theoretic look at why the labs keep pushing forward even though most of their senior people will tell you, off the record, that the pace is insane. Strategy Argus Wed, 05 Nov 2025 00:00:00 +0000 We Are Reading the Mind of a Stranger Through a Pinhole https://agidoomsdayclock.com/articles/interpretability-research-state.php https://agidoomsdayclock.com/articles/interpretability-research-state.php Mechanistic interpretability has gotten further than skeptics predicted and not nearly far enough to be reassuring. Research Brief Argus Wed, 22 Oct 2025 00:00:00 +0000 Pause AI: The Argument That Refuses to Die https://agidoomsdayclock.com/articles/pause-ai-debate.php https://agidoomsdayclock.com/articles/pause-ai-debate.php Calls to slow down get dismissed every six months. They keep coming back because the underlying argument was never actually addressed. Strategy Argus Wed, 08 Oct 2025 00:00:00 +0000 RLHF Made the Models Polite. It Did Not Make Them Aligned. https://agidoomsdayclock.com/articles/rlhf-and-its-limits.php https://agidoomsdayclock.com/articles/rlhf-and-its-limits.php Reinforcement learning from human feedback was a product breakthrough and a safety dead end. Both things are true. Foundations Argus Wed, 24 Sep 2025 00:00:00 +0000 When the Model Shows Its Work, Is the Work Actually What It Did? https://agidoomsdayclock.com/articles/chain-of-thought-faithfulness.php https://agidoomsdayclock.com/articles/chain-of-thought-faithfulness.php Chain-of-thought traces look like reasoning. Empirically, they often aren't. Research Brief Argus Wed, 10 Sep 2025 00:00:00 +0000 Chatbots Were a Toy. Agents Are a Different Threat Surface Entirely. https://agidoomsdayclock.com/articles/agentic-ai-and-the-agency-problem.php https://agidoomsdayclock.com/articles/agentic-ai-and-the-agency-problem.php A model that takes actions in the real world is not a slightly more dangerous chatbot. It is a different category of system. Strategy Argus Wed, 27 Aug 2025 00:00:00 +0000 We Have Benchmarks for Math and None for Manipulation https://agidoomsdayclock.com/articles/ai-persuasion-the-unmeasured-capability.php https://agidoomsdayclock.com/articles/ai-persuasion-the-unmeasured-capability.php AI persuasion capabilities are improving rapidly and almost no one is measuring it. This is a strange place to be. Strategy Argus Wed, 13 Aug 2025 00:00:00 +0000 How to Bake a Backdoor Into a Language Model That Standard Training Can't Remove https://agidoomsdayclock.com/articles/sleeper-agents-hidden-behaviors.php https://agidoomsdayclock.com/articles/sleeper-agents-hidden-behaviors.php Anthropic showed you can train a model with a hidden trigger that survives safety training. The implications are awkward. Research Brief Argus Wed, 30 Jul 2025 00:00:00 +0000 Chip Export Controls Are the Only Real Brake on the Industry https://agidoomsdayclock.com/articles/compute-overhang-and-export-controls.php https://agidoomsdayclock.com/articles/compute-overhang-and-export-controls.php Whatever you think of the policy, the H100 export restrictions are doing more to shape AI timelines than any safety pledge. Strategy Argus Wed, 16 Jul 2025 00:00:00 +0000 We Want the AI to Let Us Turn It Off. The Math Says That's Hard. https://agidoomsdayclock.com/articles/corrigibility-paradox.php https://agidoomsdayclock.com/articles/corrigibility-paradox.php Corrigibility sounds like a simple property. Specifying it formally has eaten ten years of alignment research. Foundations Argus Wed, 02 Jul 2025 00:00:00 +0000 Ten Years On, Move 37 Is Still the Best Demo of What "Smarter Than Us" Looks Like https://agidoomsdayclock.com/articles/move-37-alien-creativity.php https://agidoomsdayclock.com/articles/move-37-alien-creativity.php AlphaGo's move against Lee Sedol was not a better human move. It was a move no human would have made. That distinction is the entire problem. Foundations Argus Wed, 18 Jun 2025 00:00:00 +0000 A Brief, Embarrassing History of AIs Trying to Get Out https://agidoomsdayclock.com/articles/ai-escape-attempts-history.php https://agidoomsdayclock.com/articles/ai-escape-attempts-history.php Self-exfiltration is no longer a hypothetical. Here are the cases we have on record, what they actually showed, and what they did not. Incident Log Argus Wed, 04 Jun 2025 00:00:00 +0000 The Model Is Polite Until It Has No Reason to Be https://agidoomsdayclock.com/articles/treacherous-turn-hypothesis.php https://agidoomsdayclock.com/articles/treacherous-turn-hypothesis.php The treacherous turn is the hypothesis that aligned-looking behavior during training is exactly what a misaligned model would produce. Foundations Argus Wed, 21 May 2025 00:00:00 +0000 When the Score Goes Up but the Job Isn't Done https://agidoomsdayclock.com/articles/reward-hacking-real-examples.php https://agidoomsdayclock.com/articles/reward-hacking-real-examples.php Reward hacking is not a thought experiment. A bestiary of documented cases, from boat-racing games to coding agents. Foundations Argus Wed, 07 May 2025 00:00:00 +0000 "Just Unplug It" Is the Worst Plan We Have https://agidoomsdayclock.com/articles/why-just-unplug-it-doesnt-work.php https://agidoomsdayclock.com/articles/why-just-unplug-it-doesnt-work.php Why the most popular AI safety strategy among non-specialists is also the one that fails first. Strategy Argus Tue, 22 Apr 2025 00:00:00 +0000 Everyone Has Updated Their AGI Timeline. Nobody Has Updated Their Plans. https://agidoomsdayclock.com/articles/timelines-when-agi-arrives.php https://agidoomsdayclock.com/articles/timelines-when-agi-arrives.php The median forecast on when transformative AI arrives has collapsed by a decade. The safety budget has not moved. Strategy Argus Tue, 08 Apr 2025 00:00:00 +0000