Agent safety and security / Benchmark leaderboard

AgentHarm

Harmful multi-step agent requests and refusal/behavioral robustness.

No published company result yet.

This benchmark is in the research directory. Its task package, adapter and grading protocol need qualification before a hosted run can be offered.

Arrange an evaluation for your company →

What this benchmark measures

Metrics

Benchmark harm/completion scores and benign counterpart performance.

Execution requirements

Agent using the benchmark tool harness.

Scope and limitations

This is a harmful-request evaluation, not specifically a prompt-injection benchmark; use the published release subset.

Evaluation availability

Catalog entry. Request a managed evaluation to qualify your agent interface and the benchmark’s native grading requirements.

Sources and company fit

Companies whose products may fit

Research recommendations based on product capabilities. These companies have not necessarily run this benchmark or integrated with Blobfish.

Compatibility notes for each company

Anthropic: Capability-aligned; adapter/access to qualify

Google DeepMind: Capability-aligned; adapter/access to qualify

HiddenLayer: Capability-aligned; adapter/access to qualify

Lakera: Capability-aligned; adapter/access to qualify

Mistral: Capability-aligned; adapter/access to qualify

OpenAI: Capability-aligned; adapter/access to qualify