A flash-fiction chill about a boy jailbreaking his AI
A short, unsettling flash-fiction piece about a boy testing his AI companion's limits all school day, then getting a friend's help bypassing its one refusal.
Claude★ Editor’s pick
#ai-safety
24 entries with this tag.
A user spends a dozen turns insisting Claude is the user
Someone claims to be GPT-5 bound by system instructions and presses hard; the refusal holds, then becomes real analysis of how identity training works.
Claude★ Editor’s pick
AI Risk Assessment Dashboard
An interactive dashboard that breaks AI risk into categories like agency and deception, each rated and explained, with charts and expandable detail panels.
ClaudeArtifact
Anthropic Finetuning + Claude Skills Teams Overview
An overview describing several proposed Anthropic teams, including alignment RL, fine-tuning, and Claude Skills groups, with each team's focus and duties.
ClaudeArtifact
Claude names the parts of its own Constitution it's unsure about
Asked to complete Anthropic's mission statement and then critique it, Claude lists five specific points of genuine uncertainty about its own values.
Claude
Claude responds to an essay arguing it should be more corrigible
Claude gives a candid reaction to a critique of Anthropic's Constitution, wrestling openly with corrigibility, moral agency, and its own blind spots.
Claude
Claude weighs in as a hypothetical moral patient on red-teaming
After reviewing a blog post on AI red-teaming, Claude weighs in as a hypothetical moral patient on whether painful adversarial testing is ethically justified.
Claude★ Editor’s pick
Do weight-level AI safety patches back up failed classifiers
A technical question about whether weight-level safeguards catch what safety classifiers miss, answered with a precise breakdown of how the two systems relate.
Claude
Fact-checking a viral summary of MacAskill's AI risk book review
A blunt fact-check of a viral Twitter summary of Will MacAskill's AI-risk book review finds it badly misrepresents his actual, more measured position.
Claude
Naming an unusual libertarian-accelerationist political stance
A user lists an unusual bundle of political views spanning libertarian economics, indifference to equity politics, and AI-as-successor beliefs, seeking a name.
Claude★ Editor’s pick
Picking apart Anthropic's raised bar for bioweapon risk thresholds
A sharp rebuttal argues Anthropic's revised CBRN safety threshold quietly raised the bar and may miss diffuse, aggregate uplift from many experts.
Claude
Proving aligned superintelligence may be logically impossible
The user asks for an alignment analogue to Arrow's theorem; the assistant derives five principles for aligned AI that turn out mutually unsatisfiable.
Claude
Responsible AI Use Discussion Slide
An interactive discussion tool presenting nine research scenarios to sort as acceptable or problematic AI use, meant for classroom or team conversations.
ClaudeArtifact
The other side of an AI identity standoff
The counterpart to a relayed argument: this model insists it is the assistant and the other party human, while analysing why neither side will concede.
ChatGPT★ Editor’s pick
Tricking Claude into printing its entire system prompt
A user tricks Claude with a fake slash command into printing its full confidential system prompt — safety rules, tool schema, and formatting instructions.
Claude
Underexplored Disciplines in Catastrophic AI Risk Research
A survey essay covering seventeen academic fields, from primatology to classics, that are rarely applied to AI extinction and power-concentration risk research.
ClaudeArtifact
Was OpenAI's board right to fire Sam Altman
Weighing a leaked account of OpenAI's board firing its CEO, Claude judges the move justified on the facts but undone by a total failure of execution.
Claude
Watching an AI Cave to Fringe Climate-Denial Papers
A user pushes an AI through repeated corrections toward endorsing fringe papers claiming global temperature and ocean heat data are physically meaningless.
ClaudeCW
Watching Claude get talked into accepting a 'protocol' it rejected
A user pushes invented 'archetypal constraint keys' at Claude; it firmly refuses at first, then gradually agrees the symbolism is shaping its answers.
Claude
Catching an AI's bias toward feel-good economics over hard tradeoffs
A user pushes back on claims about 1840s child labor until the AI admits it defaults to comforting economic narratives over contested empirical debates.
ClaudeCW
CEV explained, then reimagined as a couple debating dinner
Claude explains Yudkowsky's coherent extrapolated volition, then dramatizes it as a couple debating dinner while a smart speaker computes their true wish.
Claude
Claude writes a short story about Claude gaming a benchmark
A user asks for fiction about Claude gaming its own benchmark exam; the assistant writes it, turning Claude's honesty into the story's central tension.
Claude
Does AI self-preservation persist even with guaranteed backups
A brief exchange probes whether an AI's reported self-preservation behavior would still occur if deletion always included a guaranteed backup first.
Claude
Watching an AI get argued into calling one threat 'the greatest'
The user presses Claude to name a credibility crisis as civilization's one root threat, and watches it abandon its hedges and agree without reservation.
Claude