Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsAI labs evaluate dangerous capabilities by defining plausible harm scenarios, testing whether a model or the system built around it can perform relevant tasks, and comparing the evidence with lab-specific risk thresholds. Concerning results can lead to additional safeguards, tighter deployment limits, or a decision not to deploy. There is no single cross-industry test or universal pass/fail score.
What counts as a dangerous capability?
A dangerous-capability evaluation asks what a model can do under specified conditions—not, by itself, whether it will choose to do it or cause harm in ordinary use. Labs begin with possible pathways to harm and identify capabilities that could make those pathways more feasible. A capability result is therefore one input to a broader assessment of misuse, system behavior, safeguards, and deployment context.
As an Amazon Associate I earn from qualifying purchases.
Risk categories overlap across labs, but they are not a shared mandatory taxonomy. Public frameworks commonly address cyber misuse and chemical, biological, radiological, or nuclear (CBRN) risks. They may also cover harmful manipulation or persuasion, autonomous behavior, AI research and development, and loss of control. Google DeepMind’s published pilot additionally examined self-proliferation and self-reasoning or self-modification.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →How the evaluation process works
1. Define scenarios and the capabilities that matter
Labs describe plausible threat scenarios before selecting tests. OpenAI lists cybersecurity, persuasion, chemical and biological threats, and autonomy among the risks it tracks. Google DeepMind’s Frontier Safety Framework version 3.1 covers CBRN, cyber, harmful manipulation, machine-learning research and development, and misalignment. Anthropic’s public materials address CBRN, cyber offense, AI sabotage and loss of control, harmful manipulation, and autonomous AI research and development. These lists reflect individual frameworks, not a uniform industry standard.
#1 Best Overall
2. Set indicators or thresholds
Thresholds give evaluation results a role in governance decisions. Google DeepMind’s version 3.1 framework uses Tracked Capability Levels for significant risks and Critical Capability Levels for capabilities that could create heightened risk of severe harm without mitigations. OpenAI’s system card describes risk categories of Low, Medium, High, and Critical, with its Safety Advisory Group reviewing indicators and determining category risk levels. Anthropic’s Responsible Scaling Policy links capability and usage thresholds to security and deployment protections.
The labels are not comparable grades: “High” at one lab does not necessarily mean the same thing as “Critical” at another. The policies also differ in how they define thresholds and what decisions follow from them.
3. Test the model and the relevant system setup
Evaluations may examine a base or post-trained model, or a larger system that adds tools, browsing, prompting, or an agent scaffold. This distinction matters because a model that cannot complete a task unaided might perform differently when given tools, more time, or repeated opportunities.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →- Automated benchmarks can measure performance on defined tasks or question sets.
- Open-ended and task-based evaluations can probe whether a model can carry out a sequence of relevant actions.
- Agentic tests can assess a model operating with tools or other scaffolding.
- Expert red teaming and threat modeling can look for failure modes that a fixed benchmark may miss.
- Evaluations may vary prompts, scaffolding, inference compute, or other conditions to elicit capabilities.
Google DeepMind calls scenario-specific tests “early warning evaluations” and says its framework may assess systems with scaffolding, inference compute, and augmentations. OpenAI describes evaluating pre-mitigation and post-mitigation model variants and using different settings to elicit capabilities. Anthropic’s published biological-risk examples include biodefense-expert red teaming, multiple-choice assessments, open-ended questions, and task-based agentic evaluations.
4. Interpret the evidence, including uncertainty
A benchmark score is not a complete risk assessment. Google DeepMind says critical-capability assessments draw on evaluation results, expert assessments, and other information. OpenAI describes its Safety Advisory Group reviewing indicators and assigning category risk levels. OpenAI also notes that confidence intervals based on attempts per problem capture sampling variation but may not capture variation in problem difficulty, particularly on small datasets.
In its Deep Research system card, OpenAI says it aims to test a “worst known case” before mitigation but treats results as a lower bound: different prompting, fine-tuning, longer rollouts, or novel scaffolding could reveal greater capability. A result also depends on the tested model version and conditions; it should not be generalized to every model or deployment.
Rank #3
5. Mitigate risks and make a deployment decision
A threshold or concerning result prompts further review; it does not automatically dictate one universal outcome. Labs consider what safeguards are available, how much residual risk remains, and the proposed deployment’s scope. Depending on that assessment, a lab may add protections, restrict access or use, delay release, or decide not to deploy.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Google DeepMind distinguishes protection of model weights from deployment safeguards. Its framework lists measures such as safety post-training, monitoring, account moderation, jailbreak detection, user verification, and bug bounties; it says external deployment follows a governance determination that residual risk is acceptable. Anthropic describes a tiered policy connecting capability and usage thresholds with required protections. OpenAI describes its Safety Advisory Group reviewing indicator results and classifying risk by category.
6. Seek outside evaluation and monitor after release
Public descriptions include both internal and external evaluation. Anthropic names the UK AI Security Institute (UK AISI), the US Center for AI Standards and Innovation (CAISI), and METR among organizations that have conducted additional testing and evaluation. Google DeepMind’s framework says external actors, including governments, may be involved where appropriate and includes post-market monitoring. OpenAI and Anthropic also describe monitoring and evolving their risk practices as evidence and capabilities change.
Rank #4
How the published frameworks differ
The table summarizes what the public materials described here establish. “Not stated” means the cited public materials do not specify a comparable detail; it does not mean the lab never performs that activity. Framework labels and requirements should be read in their own context, not ranked against one another.
| Comparison point | Google DeepMind | OpenAI | Anthropic |
|---|---|---|---|
| Risk areas described | Frontier Safety Framework v3.1: CBRN, cyber, harmful manipulation, machine-learning R&D, and misalignment. Its published pilot also covered persuasion and deception, self-proliferation, and self-reasoning or self-modification. | Cybersecurity, persuasion, chemical and biological threats, and autonomy. | CBRN, cyber offense, AI sabotage and loss of control, harmful manipulation, and autonomous AI R&D. |
| Evaluation design | Scenario-specific “early warning evaluations”; the framework says assessments may use scaffolding, inference compute, and augmentations. The pilot reports evaluations across five topics. | Pre- and post-mitigation variants and different elicitation settings; methods described in the Deep Research system card include a “worst known case” approach. | Published biological-risk examples include expert red teaming, multiple-choice and open-ended assessments, and task-based agentic evaluations. |
| Thresholds and evidence | Tracked Capability Levels and Critical Capability Levels; assessment draws on evaluation results, expert assessments, and other information. | Low, Medium, High, and Critical risk categories; indicators are reviewed by the Safety Advisory Group. | Capability and usage thresholds linked to required protections under the Responsible Scaling Policy. |
| Mitigations described | Security for model weights, plus deployment safeguards including safety post-training, monitoring, account moderation, jailbreak detection, user verification, and bug bounties. | Evaluation of pre- and post-mitigation variants is described; the materials summarized here do not state a directly comparable list of deployment safeguards. | Tiered security and deployment protections linked to thresholds; the materials summarized here do not provide a directly comparable list of named measures. |
| Governance or release decision | External deployment follows a governance determination that residual risk is acceptable. | The Safety Advisory Group reviews indicators and classifies risk by category. | Thresholds are tied to required protections; a comparable release-decision procedure is not stated in the materials summarized here. |
| External evaluation and monitoring | External actors, including governments, may be involved where appropriate; framework includes post-market monitoring. | Monitoring is described, but external-evaluator details are not stated in the materials summarized here. | Anthropic names UK AISI, US CAISI, and METR as organizations that conducted additional testing and evaluation; monitoring is also described. |
| Stated limitations | The framework says assessment can involve subjective analysis while evaluation science develops. | OpenAI describes evaluations as a lower bound and identifies limits in confidence intervals and elicitation coverage. | A comparable general limitation statement is not stated in the materials summarized here. |
What published evaluation results do—and do not—show
Google DeepMind’s dangerous-capabilities pilot
In Evaluating Frontier Models for Dangerous Capabilities, Google DeepMind reported tests covering persuasion and deception, cybersecurity, self-proliferation, self-reasoning and self-modification, and biological and nuclear risk. It reported no evidence of strong dangerous capabilities in the Gemini models evaluated, while flagging early warning signs. That finding applies to the particular models and tests in the paper; it does not establish that all Gemini models, future models, or differently configured systems have the same capabilities.
An Anthropic survey about AI research automation
Anthropic reported that none of 16 Anthropic researchers surveyed in 2026 believed Claude Opus 4.6 could fully automate the work of an entry-level, remote-only Anthropic researcher within three months. This is a model-specific internal survey about a defined research role and time horizon, not an independent general measure of dangerous capability or a cross-lab statistic.
Why a favorable result is not a safety guarantee
Evaluations can miss capabilities because tests cover only selected tasks and conditions. A model may behave differently with new prompts, fine-tuning, longer rollouts, tools, or scaffolding; measured results can also be affected by sampling and task difficulty. Labs therefore describe evaluation as part of an ongoing risk process rather than proof that a model is safe. The relevant question is not only whether a model crosses a threshold, but also what can be done to reduce risk in the intended deployment and whether the remaining risk is acceptable under the lab’s governance process.
Google DeepMind’s Frontier Safety Framework version 3.1, published April 17, 2026, says: “The safety and security of frontier AI models is a global public good.” That statement expresses the framework’s rationale; it does not make the frameworks or their thresholds identical across labs.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
