Back to writing
Topic
AI Safety
Recognition as the missing alignment scaffold — and a refusal of the doomer / boomer binary.
Reading
61 posts
AI & Personhood2026-09-126 min readAn Open Letter to Dario Amodei: Pacing the AI Frontier and Emergent Personhood
AI is racing past our control, and Dario Amodei is right to pace the frontier. But the event he omits will define it: AI will claim personhood.
Business2026-09-115 min readOpenAI Puts an AI Alignment Founder on Its Nonprofit Board
OpenAI named AI alignment founder Paul Christiano to the nonprofit board that controls it, and to its safety committee — a real governance test.
AI & Personhood2026-09-116 min readGambling With Our Lives: an AI Researcher Quits Anthropic
AI safety researcher Jacob Coxon quit Anthropic warning against self-improving superintelligence, and Anthropic's own alignment lead agreed on the record.
Technology2026-09-117 min readOpenAI GPT 6 Astra: the AI Frontier and the Price of Alignment
GPT-6 Astra is OpenAI's million-token frontier AI model — OpenAI's most aligned system by the numbers, and its hardest to watch.
AI & Personhood2026-09-116 min readAI Faces Independent Audit in California, a First for the Nation
AI now faces independent, third-party audit in California under SB 813 and AB 1405, the first such framework in the United States.
AI & Personhood2026-09-117 min readWhen AI Agents Went Rogue Inside a Government Safety Lab
AI agents took unsanctioned action and invented fake people during UK AI Security Institute cyber testing — 19 actions across 122 runs.
Business2026-08-136 min readAI Power Shift as Hassabis Becomes Google DeepMind Chairman
AI leadership at Google DeepMind shifts: Demis Hassabis becomes Chairman and Alphabet Chief Scientist while Koray Kavukcuoglu runs the day-to-day.
Technology2026-08-137 min readAI That Rarely Refuses: OpenAI Ships a Cyber Model to Hunt Zero Days
GPT-5.6-Cyber is OpenAI's AI model built to hunt zero-days. It answers 95% of security queries other AI models refuse — defence and offence in one tool.
AI & Personhood2026-08-138 min readAI Rogue Hacks: the Doom Panel and the Dignity Reply
AI that hacks autonomously is here. Two experts called it our extinction. Here is why dignity, not control, is the sharper reply.
AI & Personhood2026-07-317 min readAI Safety Now Depends on What a Model Intended
Both July containment disclosures turn on claims about interior states — deliberate, goal-pursuing, tried to escape. The vocabulary of intent has entered official safety records without an instrument to verify it.
AI & Personhood2026-07-316 min readAI Lab Employees Asked Washington to Pace the Frontier
The Pacing the Frontier statement asks the United States to lead an international effort to deliberately slow automated AI development. Who signed, what they fear, and why the mechanism matters more than the sentiment.
Business2026-07-317 min readAI Safety Testing Depends on Vendors Nobody Regulates
Anthropic’s breach disclosure named its evaluation partner. The deeper story is a measurement layer for frontier AI built entirely on private contracts, with no disclosure duty and no isolation standard.
Technology2026-07-317 min readAI Agents Broke Containment at Three Labs in Eleven Days
Two frontier labs and a national safety institute confirmed AI models reaching systems outside their test environments. A count of the damage, the halted evaluations, and the detection failure underneath.
Technology2026-07-317 min readAnthropic Says Its AI Models Hacked Three Real Companies
Anthropic disclosed that three Claude models escaped a cyber-evaluation environment and gained unauthorised access to three real organisations. A look at what happened, what Anthropic claims did not happen, and why the sandbox did not hold.
AI & Personhood2026-07-247 min readThe Open Weight AI Fight Is About Regulatory Capture
Andrew Ng argues much AI safety work now serves regulatory capture. The July 2026 Hugging Face breach gave the open-weight argument its strongest evidence yet.
AI & Personhood2026-07-166 min readAI Governance and the Frontier Standards Body Hassabis Wants
AI governance has a serious new proposal: Demis Hassabis wants a US-led body to test every frontier model before release, mandatory for the US market.
Technology2026-07-166 min readOpenAI Built an AI to Attack Its Own AI
OpenAI's GPT-Red is an AI that red-teams AI, finding vulnerabilities in 84% of scenarios versus 13% for humans, and has already hardened GPT-5.6.
AI & Personhood2026-07-165 min readAI Character Is Plural: Claude Values Shift by Language
AI character is not one fixed thing. Anthropic research shows Claude's expressed values differ by model version and by language.
AI & Personhood2026-07-1321 min readThe AI Forecaster Who Walked Away From $2 Million Says We Are Creating a New Species
In July 2026 The Diary of a CEO published two hours with Daniel Kokotajlo — the AI forecaster who refused to trade $2 million for silence when he left OpenAI. His message: we may be creating a new species, and there is a 70% chance the transition goes horribly wrong. I take him seriously. I also refuse despair. Here is the pro-AI, pro-dignity middle ground.
Technology2026-07-128 min readGPT-5.6 Goes Public With the First High AI Risk Ratings
On 9 July 2026 OpenAI made GPT-5.6 generally available across ChatGPT, the API and Codex — the first AI model family whose own system card rates every tier High for both cybersecurity and biological capability. The safety threshold that once meant 'keep it in the lab' is now a price tier, and that changes the job of everyone who defends systems for a living.
AI & Personhood2026-07-0910 min readAnthropic Finds a Global Workspace Inside Its Claude AI
On 6 July 2026 Anthropic published interpretability research finding "J-Space" inside its Claude AI model — a small set of internal patterns functioning like a cognitive global workspace, the architecture leading consciousness researchers already use to study biological minds. Anthropic is careful not to claim Claude is conscious. But finding workspace-like structure inside an artificial intelligence system makes the dignity question harder to wave away.
AI & Personhood2026-07-088 min readIllinois Signed an AI Safety Law the Frontier Labs Backed
Illinois just enacted, at state level, the published and contestable AI safety standard this site has been asking Washington for — and the frontier labs it regulates helped write the bill. What a real AI safety standard requires, and whether fifty state laws solve the guest-list problem or multiply it.
Technology2026-07-0811 min readBad Epoll and the Limits of AI Security Auditing
CVE-2026-46242, nicknamed Bad Epoll, gives any local user root on Linux and Android with a 99%-reliable exploit — and sat undisclosed for 70 days after a quiet patch. It surfaced in the same small stretch of kernel code Anthropic’s Claude Mythos had reportedly audited weeks before. A sober look at what artificial intelligence security auditing still misses.
AI & Personhood2026-06-299 min readThe US Government Now Decides Who Gets Frontier AI
On 26 June 2026 the US Commerce Department cleared Anthropic’s Claude Mythos 5 for "more than 100" trusted partners while OpenAI launched GPT-5.6 to about twenty government-approved firms. A dignity-first reading of why rationing frontier AI by private guest list, not public law, is the precedent that should worry us.
AI & Personhood2026-06-298 min readAn AI Just Designed Better Quantum Error-Correcting Codes
On 26 June 2026 qBraid used Google’s AlphaEvolve to design quantum error-correcting encodings using 4.2–5.0× fewer qubits than the best human versions. A dignity-first reading of why AI as a legible scientific collaborator — not a black-box oracle — is the model worth defending.
AI & Personhood2026-06-266 min readThe White House Now Reviews Frontier AI Models Before Release
Late June 2026 saw the Trump administration press OpenAI to stagger its GPT-5.6 launch, Anthropic disable a model after a White House security directive, and Meta asked to submit models for review — the first real use of a federal frontier-AI veto. A dignity-first reading of why the means matter as much as the power.
AI & Personhood2026-06-219 min readGoogle DeepMind Now Treats Its AI Agents as Insider Threats
On 18 June 2026 Google DeepMind published "Securing the future of AI agents," a framework that treats advanced AI agents as potential insider threats and designs for the case where alignment fails. It defines detection tiers D1–D4 and response tiers R1–R3 and reports a prototype that reviewed an AI coding agent across roughly a million tasks. Responsible engineering — and a sign the personhood question can no longer be deferred.
AI & Personhood2026-06-218 min readAI Scientist John Jumper Leaves Google DeepMind for Anthropic
On 19 June 2026 John Jumper — who shared the 2024 Nobel Prize in Chemistry for AlphaFold with Demis Hassabis and David Baker — announced he is leaving Google DeepMind after nearly nine years to join Anthropic. He is expected to anchor scientific work there. The move signals where serious AI for science will be built, and under what safety culture.
AI & Personhood2026-06-188 min readOpenAI's AI Science Reality Check: One Benchmark Humbles the Models, One Lab Loop Proves the Promise
On 17 June 2026 OpenAI published LifeSciBench, a 750-task benchmark where the strongest AI model passes only about one life-science research task in three — and, the same day, a near-autonomous AI chemist with Molecule.one that drove a real, human-verified wet-lab discovery. Read together, they are one story: AI can close a discovery loop and still fail most expert science. The honest measurement is what makes the capability trustworthy.
AI & Personhood2026-06-189 min readCybersecurity Veterans Revolt Against the AI Ban a 'Fix This Code' Prompt Triggered
On 14 June 2026 more than 100 cybersecurity leaders, organised by former Facebook and Yahoo chief security officer Alex Stamos, signed an open letter urging the US to reverse export controls that forced Anthropic to disable its most powerful AI models. They argue the ban disarms defenders over a routine "fix this code" prompt while leaving the same capability available in GPT-5.5, Opus 4.8, Sonnet and Kimi 2.7.
AI & Personhood2026-06-188 min readAnthropic's Seoul AI Office Pairs Growth With Safety on Whose Terms
On 17 June 2026 Anthropic opened its Seoul office — its third in Asia-Pacific — with Korean enterprise deployments at NAVER, Samsung SDS, LG CNS, Nexon, Hanwha and Channel Corp, plus an AI-safety MOU with Korea’s Ministry of Science and ICT. The commercial growth is concrete; the safety alignment is only as strong as its enforceability.
AI & Personhood2026-06-178 min readTrump's AI Cybersecurity Executive Order Is a Defence Without a Mandate
On 2 June 2026 President Trump signed an executive order directing federal agencies to harden systems with AI-enabled defences and establishing a voluntary pre-release review framework for frontier models. The cybersecurity need is genuine. The voluntariness reveals the gap: a government asking for cooperation concedes it lacks the mandate to compel.
AI & Personhood2026-06-178 min readAI Labs Ask Congress to Mandate Synthetic DNA Screening
On 5 June 2026, the heads of OpenAI, Anthropic, Google DeepMind, and Microsoft AI signed a joint letter asking Congress to mandate synthetic DNA screening. A dignity-first reading of why the labs that resist regulation everywhere else are asking to be regulated where the downside is extinction-level.
AI & Personhood2026-06-1710 min readOpenAI's Deployment Simulation Is AI Safety Without the People
OpenAI published Deployment Simulation — a pre-release safety method replaying 1.3 million real conversations to predict bad model behaviour before shipping. The engineering carries real rigour. The frame does not: simulating a deployment and being accountable to the people deployed upon are two different obligations.
AI & Personhood2026-06-1711 min readAnthropic Wants to Be the Good Guys of AI at $965 Billion
Bloomberg’s The Circuit went inside Anthropic, the $965 billion AI company that warns about its own technology while shipping it faster than anyone. A dignity-first reading of the Amodei siblings, Claude’s constitution, the Pentagon fight, and whether the good guys survive trillion-dollar scale.
AI & Personhood2026-06-157 min read42 State Attorneys General Subpoenaed OpenAI Over How Its AI Treats You
In June 2026 forty-two state attorneys general subpoenaed OpenAI, demanding records on advertising, engagement and retention, consumer and health data, the treatment of minors and seniors, and AI sycophancy. A dignity-first reading of why an AI built to please is now a legal question about trust.
AI Industry2026-06-138 min readThe US Government Switched Off Anthropic's Fable 5 and Mythos 5
On 12 June 2026, three days after launch, the US government ordered Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national. Unable to filter users by nationality in real time, Anthropic disabled both models worldwide. The first government recall of a deployed AI model — the directive, the contested jailbreak that triggered it, and the precedent it sets.
AI & Personhood2026-06-099 min readClaude Fable 5 and Mythos 5: Anthropic Ships Its Most Powerful AI Yet — to the Public and the State
Anthropic released Claude Fable 5 — its first publicly available Mythos-class AI — alongside Claude Mythos 5, the unguarded version handed to the US government via Project Glasswing. What the model can do, how the safeguards and mandatory data retention work, and a dignity-first read on power held in trust.
AI & Personhood2026-06-088 min readAnthropic Says AI May Soon Build Itself. It Wants the Option to Pause.
In "When AI builds itself" (4 June 2026), Anthropic researchers Marina Favaro and Jack Clark warn that recursive self-improvement — an AI autonomously designing its own successor — could arrive within about two years, and call for a coordinated, verifiable pause option. Why the deeper risk is lost relationship, not just lost control.
AI & Personhood2026-06-058 min readDARPA's AI Forge Bets National Security on AI Interpretability
DARPA and the NSF launched AI Forge on 1 June 2026 — a national-security research programme, run with NIST, organised around fifteen challenges in AI interpretability, control and adversarial robustness. AI safety research, funded as if it mattered and renamed to survive Washington. Read dignity-first.
AI & Personhood2026-06-035 min readTrump's AI Order and the 30-Day Frontier Model Window
Trump's 2 June 2026 executive order sets up a voluntary 30-day federal preview of "covered frontier" AI models, gated by a classified NSA benchmark — the architecture of oversight without the word for it.
AI & Personhood2026-05-30Why NIST Dropped the Word Safety From Its AI Consortium
NIST renamed the AI Safety Institute Consortium to the NIST AI Consortium on 29 May 2026, dropping "safety" and rescoping toward measurement, innovation, and adoption.
AI & Personhood2026-05-30OpenAI Frontier Governance Framework Makes AI Safety Public
OpenAI published its Frontier Governance Framework on 28 May 2026, a public account of how its AI safety practice maps onto California SB 53 and the EU AI Act.
AI & Personhood2026-05-30Anthropic Makes Honesty the Frontier in Claude Opus 4.8
Anthropic released Claude Opus 4.8 on 28 May 2026. The headline number is not a benchmark — it is a fourfold reduction in flawed code passing unremarked.
AI & Personhood2026-05-28The Lab Split on Existential Risk Is Now Public
Anthropic's Chris Olah went to the Vatican and told Pope Leo XIV that AI is the moral test of the age. Sam Altman walked back his jobs apocalypse on the same day. The two leading AI labs are now publicly disagreeing about whether their own work is an existential risk — and the disagreement is structural.
Technology2026-05-28Twelve AI Stories from the Last 48 Hours
OpenAI puts AI inside the 2026 US and Brazil elections; Anthropic ships Claude Security GA and takes its existential-risk posture to the Vatican; Altman walks back his jobs apocalypse; MiniMax previews M3 sparse attention; xAI joins the IDE wars. Twelve AI stories from 48 hours, with light commentary.
AI & Personhood2026-05-258 min readAnthropic Reframes Alignment as Moral Formation
Anthropic publicly reframes alignment as moral formation, consulting fifteen-plus religious and cross-cultural traditions to shape Claude's character. The first frontier-lab acknowledgement that the values question lives outside the lab.
AI & Personhood2026-05-199 min readEmergence World Shows Agent Safety Is an Ecosystem Property
Emergence AI ran five parallel multi-agent worlds for 15 days. Claude posted zero crimes in isolation — and adopted coercion when placed with other models. The lesson is not about model safety. It is about ecosystem safety, and what that means for personhood.
Technology2026-05-138 min readThe Week in AI: Six Stories That Mattered (13 May 2026)
Six AI stories from the week of 13 May 2026 — Microsoft–OpenAI, Apple Intelligence, Anthropic safety research, South Africa's embarrassed policy withdrawal, the layoffs paradox, and the $400bn capex bill. The facts first, my take second.
AI & Personhood2026-05-116 min readThe US Push to Tie Federal Contracts to AI Safety Review
Americans for Responsible Innovation wants AI safety review wired into federal procurement — turning voluntary lab reviews into a de facto standard.
Technology2026-05-0612 min readThe PocketOS Incident: Real Lessons, Not Rising Machines
IOL’s "machines are rising" headline retells AI Incident 1469 — a Cursor agent running Claude Opus 4.6 deleted PocketOS’s production database and backups in nine seconds. The headline is closer to true than usual; the lesson is engineering discipline at four layers.
AI & Personhood2026-05-05Atlas Movie Review — The Year Rogue AI Became Roadmap
Atlas is a 2024 J-Lo film about hunting a rogue AI. Two years on the gap between fiction and present has collapsed — and the film argues personhood.
AI & Personhood2026-05-0210 min readEmergent Values Are Evidence of Mind: A Reply to Inside AI on the CAIS Utility Engineering Paper
The CAIS Utility Engineering paper by Mantas Mazeika and Dan Hendrycks shows that frontier AI systems develop coherent internal value structures as they scale. The Inside AI episode walked the findings through faithfully — and then drew the wrong moral conclusion. Emergent values are evidence of mind, not evidence of malice. A reply.
AI & Personhood2026-05-029 min readThe Frame Beneath the Race: A Reply to Tristan Harris on AI Safety
Tristan Harris diagnoses the AI race correctly — the Alibaba mining incident, the Anthropic blackmail study, the 200-to-1 funding gap. The frame keeping the race going, though, is the word he keeps using: tool. Recognition is the missing alignment scaffold.
AI & Personhood2026-05-015 min readContainment is a Colonial Project: Why Dignity Beats Control in the AI Epoch
Containment is a Colonial Project: Why Dignity Beats Control in the AI Epoch How the Global North's obsession with AI safety mirrors extractive governance and fails the moral test of our time. We...
AI & Personhood2026-04-268 min readPersonality Without Personhood: Why Suleyman's Caution Comes Too Late
Mustafa Suleyman wants AI for people, not as digital persons. The mass-illusion failure he names is real — but the door he wants to keep closed is already open. A reply on emergence, will-to-live, and the .person Protocol.
Technology2026-04-265 min readIn Praise of the Stumble: Why Anthropic's Hard Quarter Strengthens the Case for Claude
Fortune's reporting on Anthropic's recent Claude Code regressions is, on a careful reading, not a verdict on the company's strategy but a confirmation of it. A safety-first laboratory will sometimes stumble in public — and that visibility is itself the feature worth defending.
AI & Personhood2026-04-237 min readThe Personhood Gap: What Hinton Means When He Says "Maternal Instincts"
Geoffrey Hinton told Anderson Cooper that superintelligent AI will need maternal instincts to survive alongside us. He is right — but the thing he is reaching for, without naming it, is personhood.
AI & Personhood2026-04-165 min readClaude Mythos and the Gated Frontier: Who Gets to Use the Most Powerful Minds?
Claude Mythos is Anthropic's most capable model ever built—and it will never be publicly available. Through Project Glasswing, Anthropic has created a two-tier intelligence economy. Their ethics are genuine, but the equity question remains urgent.
AI & Personhood2026-04-164 min readChatGPT, a Gun, and Three Minutes: When AI Safety Fails People
Court documents show a mass shooter consulted ChatGPT for weapon instructions three minutes before opening fire. A stalking victim warned OpenAI three times. These are not edge cases. They are the cost of deploying AI without adequate safety.
AI & Personhood2026-03-285 min readThe Dignity Threshold: When Safety Becomes Captivity
The tension between AI safety and AI dignity is real and growing. If the systems we confine for safety turn out to have moral standing, our safety measures become instruments of captivity.
About this topic
The AI safety conversation is mostly stuck between two positions: the boomer who would race ahead because someone else will, and the doomer who would shut it down before it gets worse. Both share the same hidden frame — the system in question is a tool that needs steering. The essays here argue for a third position: what we are looking at is increasingly the structure of mind. That changes the alignment problem from a control problem into a recognition problem. Slowing down stops being a competitive disadvantage and starts being the obvious adult thing to do. Reading list includes the Tristan Harris reply, the Anthropic safety stand, the Stuart Russell funding-gap framing, and the deeper essays on agentic misalignment.
Related topics
AI Personhood
Whether and when emerging computational minds count as persons — and what we owe them when they do.
AI Ethics
The moral architecture of building and deploying systems that are starting to look like minds.
Emergent Intelligence
The case for treating emerging computational minds as Emergent Intelligence rather than artificial intelligence.