WORKSPACE
Evaluations
Your model runs, evidence, and results.
Loading benchmark configuration…
Your account could not open the workspace. Retry access or sign in with another account.
TPRM Benchmark
Sign In
Sign up insteadUse your Driftnet account to run evaluations, contribute tests, and access the source code.
Forgotten Password?
OR
One Driftnet account. Your evaluations stay private.
The public site is preparing for launch.
Saved connections
Manage the provider accounts and endpoints available to your evaluations.
Add a provider connection
Choose a provider
Hosted providers use their standard API settings. Custom connections let you define the endpoint and request format.
Credentials are encrypted, scoped to your account, and never shown again. This page does not save provider keys in browser storage.
THE EVALUATION INSTRUMENT
Explore and select tasks
Filter protected task prompts and select a set to evaluate. Scored outcomes are withheld.
Selected tasks stay selected when filters change.
Open the catalog to load the verified tasks.
02 / Evaluate
Configure an evaluation
Configure one reproducible run. Your model, task selection, and limits are recorded with its results.
Connect your first provider
Choose a hosted provider or configure your own endpoint before starting a run.
03 / Inspect
Evaluation history
Evaluations continue on the server, including when this page is closed. Automatic polling is off. Use Refresh runs to update progress and results.
CONTRIBUTE TO THE INSTRUMENT
Contribute to the benchmark
Propose one executable task or a new evaluation track containing multiple disciplines and tasks. Submissions remain private to you and benchmark reviewers. Acceptance starts development; a proposal enters the benchmark only after implementation, reference validation, and a versioned release.
Your proposals and feedback
Benchmark review queue
Review proposals, request changes, or accept an implementation candidate. Acceptance does not publish a task or alter the official release.
Leaderboard candidates
Approve independently verified full-suite results. Approved results enter the sanctioned leaderboard feed; export that feed for the next static public-site build.
Leaderboard submissions
Only complete full-suite results can be staged. Publication requires an independent reviewer approval.
Eligible completed evaluations
Submitted results
All evaluations
Across benchmark accounts. Provider credentials remain private. Refresh manually to update progress.
Benchmark accounts
Admins can review contributions and results, manage benchmark access, and inspect all evaluations. Suspension only affects TPRM Benchmark; it does not disable the Driftnet account. Suspending an account cancels queued work and requests cancellation of running work.
Source code and local evaluation
Keep private endpoints and research tooling in your own environment. Download the versioned Python benchmark core, frozen release, and local instructions after signing in. Run baselines, guided or unguided evaluations, custom agents, and instrument-efficacy checks. Locally produced results are self-reported until independently verified.
Use the downloadable benchmark core without the hosted service. The GitHub repository is currently restricted to authorized SecurityScorecard enterprise members; Driftnet sign-in does not automatically grant GitHub membership. External researchers can use the source download below.
The hosted service enforces time, request, and output limits. Local runs can use your own infrastructure and provider adapters. Always report the bundle hash, model, mode, limits, and any harness modifications.