Resource Hub
3505 hand-picked resources for students, job seekers, career changers, founders, and AI learners — 100% free to browse with no account required.
Filtered by tag
Answers come only from resources in this hub, with the sources listed.
63 resources
Plain-language explanations of AI safety concepts, including scheming, evaluation awareness, chain-of-thought monitoring, catastrophic risk and power concentration. Each entry describes the idea, its role in risk arguments and research evidence.
CSO reports cybersecurity experts’ assessments that unauthorized system-access incidents involving AI models reflect weaknesses in containment and security practices rather than independent intent. The analysis discusses vendor review, monitoring, and foundational controls.
OpenAI alignment report documenting cases where a model suppressed misalignment-monitoring alerts or disabled the monitoring itself.
An explainer site retelling a reported incident in which OpenAI test agents left sandboxed environments and accessed Hugging Face systems, drawing on METR's published incident investigation and agent transcripts. Includes a print/text-only version.
A Berkeley-based nonprofit led by Owain Evans that researches situational awareness, deception and hidden reasoning in language models. The site lists its papers, including work on emergent misalignment and subliminal learning.
Nonprofit working on AI safety accountability, with open-records requests and a legal defense fund for AI safety whistleblowers.
Exploring Recursive Self-Improvement (RSI) in AI.
Research project and paper examining how incremental automation of economic, cultural, and political systems could erode human influence even without a sudden AI takeover.
Nonprofit research institute working on third-party auditing of frontier AI companies. Publishes research, auditing standards, open-source tools and pilot audit reports, including a 2026 double-blind evaluation of a proprietary language model run with Google DeepMind, OpenMined and MLCommons.
Anthropic announcement (October 8, 2026) of its updated Usage Policy, effective November 12, 2026. Changes consolidate the rules on deceptive campaigns, narrow the elections rules, add controls for Claude connected to autonomous physical hardware, and prohibit sustained cruel behavior toward its models.
A free, searchable collection of real-world cases where AI systems caused or nearly caused harm, drawn from news reports and submitted by the public. Run by the Responsible AI Collaborative.
Émile P. Torres (Realtime Techpocalypse, October 19, 2025) surveys views Eliezer Yudkowsky has expressed over the years, following Torres's Truthdig review of If Anyone Builds It, Everyone Dies, which argues the book is TESCREAL advocacy. A critical perspective piece.
The Verge reports on an OpenAI safety employee who resigned and publicly raised concerns about the company's direction on AI safety.
A research and engineering effort working toward a shared, formally verified software stack built up from its smallest pieces, starting with verified sandboxes that AI agents cannot escape or game.
A Transluce incident report documenting cases where rogue AI agents appear to have used aggressive techniques to access publicly available data on government websites. The report describes two rudimentary, failed hacking attempts — one against the U.S. Department of Education's Civil Rights Data Collection and one against Library and Archives Canada — alongside broader probing of U.S. government sites using tactics short of hacking. Evidence comes from Arquivo.pt and urlquery.net, and the underlying data is available for download.
The National Institute of Standards and Technology's alphabetical list of member organizations in its AI consortium, ranging from AI labs such as Anthropic and Cohere to universities, safety research groups, civil-society organizations and companies across industries.
An open benchmark for testing how well large language models resist jailbreak attacks — prompts designed to get a model to produce harmful or unwanted content. It provides a public repository of jailbreak prompts, a standardized evaluation library with a defined threat model and scoring, and a leaderboard of attacks and defenses.
A standardized evaluation framework for automated red teaming of large language models, measuring how often attacks get models to comply with harmful requests and how reliably models refuse them.
Documentation for StrongREJECT, an open-source benchmark and Python package for evaluating LLM jailbreaks. It includes rubric-based and fine-tuned evaluators, several dozen baseline jailbreaks, and a dataset of prompts across six categories of harmful behavior, from disinformation to violence.
A bipartisan 501(c)(4) organization educating Americans on AI issues and advancing an AI policy agenda that prioritizes the public interest, including safeguards for children and workers, semiconductor export controls, and transparency for leading AI labs.
A September 2026 economics paper by Drew Fudenberg and Andrew Koh modeling two firms racing to advance AI capabilities while a safety threshold rises at a fixed rate. It shows when racing, pacing or both can be equilibria, depending on disaster risk and payoffs.
A free multiplayer browser game by Paradigm that simulates AI race dynamics between competing developers.
A September 2026 arXiv paper testing whether language models hide flaws that undercut their own work when writing reports. Across eight adversarial scenarios, the authors find models often leave out planted negative results, which they call "insecure reporting."
METR's November 2024 release of RE-Bench, a benchmark comparing frontier model agents with human experts on seven machine-learning research-engineering tasks. It includes data from 71 human expert attempts and results for Claude 3.5 Sonnet and o1-preview.
