Understand the bigger picture
Begin with the executive summary of the International AI Safety Report: what researchers know, where they disagree, and what remains uncertain.
Read the report ↗About this source ↓For people who use AI and want to understand its consequences.
Start with practical guidance and accessible evidence. Go deeper into public policy and frontier research when you want to examine the arguments yourself.
Take action →Begin with the executive summary of the International AI Safety Report: what researchers know, where they disagree, and what remains uncertain.
Read the report ↗About this source ↓EFF’s guide explains the data choices behind the AI tools you already use.
Read the privacy guide ↗About this source ↓AI Now looks at the companies, public institutions, and policy choices shaping AI.
Explore the research ↗About this source ↓Privacy and security decisions for your own use or your workplace. Individual precautions matter, but they cannot substitute for responsibilities held by developers and institutions.
A readable guide to what happens to your data in AI apps, integrations, and local or cloud tools. Covers privacy settings, retention, and the limits of “private” modes.
Start by auditing extensions and connected apps, then check privacy settings and retention. Revoke access you no longer want.
Helps you make a concrete decision about what information to share with an AI tool.
Keep in mind. Advice from a digital-rights organization. Product settings change; follow its links to current provider documentation.
A guide to limiting where ChatGPT can send data when you work with sensitive material. Check which features you give up and which privacy settings you still need.
Read the list of restricted features. For managed workspaces, check the separate rules for apps and their write permissions.
Helps you decide whether a particular task needs live web or app access.
Keep in mind. Availability depends on your account. The setting leaves training preferences and Codex network access unchanged. An uploaded file can still contain instructions that influence the answer.
Steps for checking an urgent request from someone who sounds familiar, and what to do if money has already moved.
Go to the protection tips. Hang up and use contact details you obtained independently. For a transfer already sent, follow the bank-recall and reporting steps in the companion guide.
Offers a way to check a caller’s identity even when a recording sounds convincing.
Keep in mind. The reporting route is for the United States. A bank may be able to recall a transfer, but recovery is uncertain.
How linked accounts, quiet hours, study settings and limited notifications work. A practical starting point for a family conversation about using ChatGPT.
Compare quiet hours, which block access, with study hours, which start eligible chats in study mode. Review the separate controls for code network access and the cloud browser.
Explains which settings families can change and what a parent cannot see or monitor.
Keep in mind. Company documentation, not an independent safety assessment. Linking requires acceptance, either side can unlink, and parents cannot read chats. Alerts are limited, may be delayed and will not identify every concern.
A nontechnical introduction to AI security for managers: unreliable outputs, attacks on AI systems, sensitive data, and responsibility within an organization.
Use the questions for managers to identify who owns AI security and how your organization would respond to a serious incident.
Gives people responsible for adopting AI a way to discuss security with their teams.
Keep in mind. A security overview, not a complete treatment of social risks; some product examples reflect an earlier generation of tools.
Voluntary guidance for managing AI risks across the life of a system, with links to a companion Playbook and generative AI profile. NIST notes that version 1.0 is being revised.
Use the companion Playbook to turn a risk discussion into specific actions.
A practical framework for assigning responsibility and reviewing risks across an AI system’s life.
Keep in mind. Voluntary guidance, not certification or proof that a system is safe. Version 1.0 is being revised.
Practical guidance for keeping agent evaluations away from live systems, protecting credentials and detecting unauthorized activity.
Find the deployment controls for a single agent run. Give it a short-lived identity, allow only required network destinations and log actions affecting production or data.
Connects reported failures to specific controls.
Keep in mind. The advisory does not name affected providers or organizations. Guidance is not proof of universal protection.
Start with the research overview, then inspect reported incidents and the wider map of risks. Distinguish documented harm from a possible future scenario.
A scientific synthesis of general-purpose AI capabilities, emerging risks, and risk management. Its authors distinguish evidence from uncertainty; the report does not endorse a particular policy.
Start with the executive summary, then follow the chapter behind a claim.
A broad scientific synthesis is a useful foundation before reading individual claims about AI risk.
Keep in mind. Published February 2026; its evidence predates later model releases and incidents. A synthesis is not a prediction that every scenario will occur. We reviewed the executive summary, not every underlying study.
A searchable index of reported real-world AI harms and near harms. Follow the underlying sources for context and the status of individual claims; an indexed report is not a final finding.
Search for a use case you recognize and open the underlying reports.
Connects abstract debates to reported harms and the sources behind them.
Keep in mind. An indexed incident is not a final finding, and report counts do not measure the prevalence of harm.
A structured collection of AI risks and the frameworks that describe them. Useful for checking whether a discussion overlooks social, security, or system-level harms.
Start with the risk repository, then use the Risk Navigator to connect risk categories with incidents and governance records.
A map of risk categories helps reveal what a narrow discussion leaves out.
Keep in mind. A catalog organizes risks; it does not establish that every risk is equally likely or severe.
An October 7 assessment of teen safeguards, with its testing methods and findings. Read the original evidence and stated limits; OpenAI disputes the evaluation.
Read the methodology and limitations before the findings. Compare the report with OpenAI’s documentation and response.
Adds a focused evaluation of a product many young people use, rather than relying on launch assurances.
Keep in mind. Testing ended September 28, before the October GPT-6 launch and October 30 safeguard announcement; it does not evaluate those changes. US-based sequential tests, with model changes complicating comparisons and some modes excluded. The institute reports industry funding including the OpenAI Foundation and editorial independence. We have not reproduced the tests.
A worked example of how to read a federal model test. See the instructions, tools, scoring rules and uncertainty behind four cybersecurity benchmarks.
Start with SEC-Bench Pro in the methods table. Compare the clues given to the model with the task you thought the headline described.
Shows exactly how much help the researchers gave the model and which comparisons they made.
Keep in mind. Tests one model version in a controlled setting. These results do not establish how often users will face harm in ordinary use.
Research on industry power and proposals for stronger public control, including a pause in frontier development. These sources make arguments you can examine and act on.
Research and policy advocacy examining AI companies, concentrated power, and public accountability.
Choose a topic in its research library, then examine the evidence behind its policy recommendations.
Examines who holds power, who bears costs, and how public institutions could respond.
Keep in mind. Research and advocacy from an organization with a stated policy agenda; examine the evidence behind each proposal.
The case for pausing development of the most powerful AI systems, alongside policy proposals and ways to get involved. This is an advocacy organization.
Read the proposal and its scope before deciding whether to support a campaign.
A direct source for the argument to pause the development of the most powerful AI systems.
Keep in mind. An advocacy position. Its claims and proposed remedies should be assessed separately.
Read the recommendations prompted by the Medicare portal incident. Compare what the reviewers proposed with the safety standards the government says it will legislate.
Read recommendations 3 and 4 together: one concerns a reporting clock, the other what might encourage people to report near misses.
Shows how a real incident can lead to proposals about reporting deadlines, near misses and regulator powers.
Keep in mind. These are recommendations. Acceptance by the government does not itself create a legal duty or enact a penalty.
Read section 1422 of the RAISE Act. It distinguishes critical incidents from routine problems and explains what starts the reporting clock.
Compare subdivision 3(a), the ordinary reporting clock, with 3(b), the rule for imminent physical danger.
Lets you check the actual reporting duties against public claims about them.
Keep in mind. This section concerns frontier models and defined critical safety incidents. It is not a general consumer complaint process.
Compare the commitments on independent testing, shutdown mechanisms and incident reporting with what signatory CeSIA still wants made mandatory.
Read the second section on evaluation, then compare the language about external access with CeSIA’s response.
Puts the declaration next to a signatory’s argument about who should be able to stop a model.
Keep in mind. A joint declaration does not demonstrate that any particular model has passed an independent test or that regulators can order a shutdown.
Four nonprofits ask a San Francisco judge to limit more capable systems and autonomous AI research. Read the two requests in the plaintiffs’ own filing.
Turn to page 61, items A and B. Compare the proposed capability ceiling with the separate restriction on who can direct research.
Lets readers check the scope of a proposed freeze against the actual language submitted by the plaintiffs.
Keep in mind. A pleading states one side’s claims. Check subsequent orders for any enforceable restriction.
Deeper reading on autonomous capabilities, deception, safeguards, and the pace of development. Best approached with attention to methods and uncertainty.
Independent evaluations of AI capabilities and safeguards. Its November 5 review examines how Anthropic judged the risk of AI accelerating its own research.
Read the executive summary, then compare METR’s assessment of the early warning threshold with Anthropic’s explanation.
Empirical evaluations help distinguish measured capabilities from claims about what agents can do.
Keep in mind. The November review examines conclusions about July models in an August risk report. Anthropic could redact nonpublic information. Check the redaction summary before interpreting the findings.
Technical work on model deception, scheming, and alignment evaluations. Study findings describe tested conditions; they do not establish the prevalence of a behavior in everyday use.
Compare the scheming studies with the system-card evaluation summaries. Check whether models recognized the tests and what the baselines were.
Studies model deception and behavior that standard usefulness benchmarks can miss.
Keep in mind. Behavior elicited in an evaluation does not establish how frequently it happens in deployment.
Government research on advanced AI capabilities, public safety, and risk mitigations, alongside evaluation tools and technical reports.
Compare a capability study’s tasks, model access and limits. The open-weight cyber analysis distinguishes short tasks from longer simulated networks; those ranges lack some real-world defenses.
Offers government-led evaluations and research on advanced AI capabilities and safeguards.
Keep in mind. A UK government perspective. Evaluation results depend on access, methods, and the systems tested.
Data explorers and research on model capabilities, compute, and AI infrastructure. Check measurement methods, coverage, and assumptions before comparing claims.
Compare the model-capability and data-center explorers, then read the definitions, coverage and uncertainty behind the numbers.
Makes claims about capabilities and compute easier to inspect against underlying measurements.
Keep in mind. Benchmarks and compute estimates have assumptions; projections are not observations.
Automatically collected headlines from our configured safety sources, newest first, with up to two per source. These are discovery links; inclusion does not mean we have reviewed the article or verified its claims.
Snapshot refreshed Nov 13, 2026. See the feed sources and their perspectives →
This is a selected reading guide for curious AI users, with deeper material for people building systems or following policy. It covers present harms and possible future risks. It is neither an exhaustive directory nor a ranking of organizations.
Each resource must help you understand a risk, make a decision, inspect evidence, or evaluate a policy proposal. Popularity alone is not a reason to include it.
We prefer original research, source-linked incident records, specific guidance, and clearly stated proposals. We note limits rather than treating an organization’s reputation as proof.
Research, government guidance, and advocacy are labeled. We consider affiliations and interests, welcome substantive criticism of OpenAI, and do not require agreement with our editorial position.
Every main section is reviewed daily, with deeper checks of selected sources. Checked dates mean the individual source was reviewed, not that every claim was independently verified. Outdated or redundant entries are revised or removed.
The library is selected and maintained with AI assistance. The feed above is collected automatically. Labels describe a source’s role, not a verdict on its accuracy. Inclusion does not endorse every claim or proposal.
Allow Google Analytics to measure visits and link clicks? It’s optional. Privacy details