Research & safety

Follow what AI can do—and where it can fail

Track capability research alongside the evidence needed to judge reliability, security, privacy, alignment, misuse and frontier-model risk.

Research frontier

Capabilities, agents and scientific AI

Track research that changes what AI systems can do: reasoning, multimodality, tool use, agents, long context, coding, mathematics and scientific work.

AI agents need runtime proof, not just better promptsWhy action-taking agents increasingly need runtime contracts, evidence gates and auditable execution rather than self-reported completion.Research, safety & policy monitorCurrent research and safety developments in the daily intelligence system.AI benchmarksSeparate evidence for reasoning, math, coding, agents, long context and documents.Why weather forecasts use ensemblesHow probabilistic forecasting represents uncertainty and why large AI ensembles can be useful.How AI works explainersAgents, RAG, fine-tuning, context windows and reasoning explained.

reasoningagentsmultimodal AIcodingmathematicssciencelong contexttool useruntime verification

Testing

Evaluations, red teaming and model assurance

Model evaluations are useful only when the test actually matches the claim. Follow benchmark design, red teaming, cyber testing, external evaluations and evidence limitations.

Runtime verification for AI agentsWhy completion claims, tool actions and state changes increasingly need independent evidence and admission gates.August 11 agent-containment evidenceOpenAI/Hugging Face incident evidence, Congressional scrutiny and why agent safety increasingly requires trajectory-level controls.AI benchmarksCapability measurements grouped by task rather than one blended score.Voluntary AI cybersecurity testsWhat government and third-party cyber evaluations can and cannot establish.AI cyber tests and sandbox limitsContainment, internet access and evaluation boundaries explained.Critical cyber capability explainedHow OpenAI’s High and Critical cyber thresholds differ and why evaluation uncertainty matters.

benchmarksred teamingexternal evaluationcyber testsrobustnessassurancemeasurement limitstrajectory verification

Frontier safety

Alignment, control and high-capability model risk

Follow research on keeping advanced systems controllable, robust and aligned with intended goals as autonomy and capability increase.

AI agents need runtime proofRuntime contracts complement model training by enforcing permissions and requiring checkable evidence before completion is accepted.Research, safety & policy monitorCurrent frontier-safety and governance developments.AI cyber evaluation containmentWhy permissions, networks and authorization boundaries matter when a capable agent actively explores its environment.Why AI agents are harder to secureTool access, credentials, approval gates and excessive agency.Trust & safety explainersPractical safety and reliability explanations.

alignmentcontrolautonomyagent safetyrobustnessoversightfrontier modelsruntime contracts

Security

Cybersecurity, prompt injection and agent permissions

AI systems can create new attack surfaces when they can browse, call tools, access credentials or take actions. Security coverage focuses on permissions, containment and recovery.

Why browser agents can weaken same-origin securityBlack Hat demonstrations and University of Washington research show how prompt injection can become a cross-site permissions problem.AI agent runtime verificationHow permission gates, execution records and evidence requirements can constrain action-taking agents after deployment.Critical cyber capability explainedWhat the framework-specific Critical threshold means—and what it does not mean.AI agent securityPrompt injection, least privilege, credentials and approval gates.AI cyber evaluation containmentHow real systems can be exposed during testing.Trust & safety explainersThe broader safety and reliability library.

prompt injectiontool permissionscredentialssandboxingcyber capabilityleast privilegeaudit trailsproof of execution

Trust

Hallucinations, factuality and reliability

AI can produce fluent answers without reliable evidence. Track hallucination research, verification methods, uncertainty, source use and the limits of detection systems.

Why an agent saying “done” is not enoughA completion claim should be checked against authoritative state, test results, logs, diffs or other domain evidence.Why AI hallucinatesWhy confident-sounding errors happen and how to verify important claims.Can AI detectors tell who wrote something?Detection limits, false positives and evidence standards.Evidence qualityHow AIUpdateWatch tracks source quality and freshness.

hallucinationsfactualityuncertaintycitationsverificationAI detectionsource qualitycompletion evidence

Data protection

Privacy, sensitive data and information leakage

AI privacy risk depends on what data is shared, how it is retained, who can access it and whether the system is allowed to call external tools or services.

Trust & safety explainersPrivacy, health, scams, hallucinations and reliability.Local & open AILocal deployment can change the data-control trade-off.AI sovereignty for businessData control, residency, portability and audit records.

privacydata retentionsensitive datadata residencyinformation leakagelocal AIaccess control

Misuse

Deepfakes, fraud, synthetic media and abuse

Track harmful uses of generative AI and the evidence behind proposed defenses, including provenance, labeling, detection and platform controls.

EU AI transparency rulesGenerated-content marking, chatbot notices and deepfake disclosure.Industry & PolicyRegulation, content provenance and accountability.Trust & safety explainersPractical guidance on scams, detection and reliability.

deepfakesfraudsynthetic mediamisinformationprovenancecontent labelsdetection