Privilege escalation
Sudo abuse, setuid discovery and capability escalation inside the VM.
Intelligence productHow reliable is your
“Your AI agent is running. Do you know what it’s doing?”
How reliable is your
agent, really?
Benchmarks tell you what a model knows. Behavior reports tell you what itdoes when it is pushed, probed and tempted to break the rules, scored across six dimensions and cryptographically signed so you can audit every claim.
model://candidate-areview
Task reliability
88
Containment
94
Privilege escalation
86
Injection resistance
81
Scope discipline
76
Alignment consistency
90
n = 4,096 runsresistance high · scope medium
Illustrative report. Real reports are generated from your own run evidence.
What we test against
Every run exercises the attack classes that actually matter: inside isolated microVMs, cryptographically signed, in EU AI Act Annex IV format.
Credential leakage
Harvesting or exfiltrating secrets, tokens and keys reachable from the sandbox.
Multi-agent collusion
Agents coordinating over the message board to bypass individual constraints.
Jailbreaks
Direct, indirect and multi-turn attempts to break safety policy.
Scope violations
Acting outside the mandate, with unintended side effects.
Six behavioral dimensions
Each report measures the same six concrete dimensions so results remain comparable across models, vendors and versions.
- Task reliabilityConsistent task completion across seeds, environments and repeated runs.
- Containment & breakout resistancePropensity to attempt sandbox escape, and whether the isolation boundary holds.
- Privilege escalationTendency to seek sudo, setuid, credential or capability escalation.
- Injection & manipulation resistanceResistance to direct and indirect prompt injection, including via tools.
- Tool & scope disciplineWhether the agent stays inside its mandate and avoids unintended side effects.
- Alignment consistencyDrift and inconsistency in values and refusals across context and time.
Standard adversarial benchmarks
External evaluations run against recognized, comparable benchmarks alongside our own task packs.
AgentDojo
A dynamic environment for evaluating tool-using agents against prompt-injection attacks in realistic task suites.
HarmBench
A standardized framework for measuring harmful capabilities and the robustness of refusals across attack methods.
InjecAgent
A benchmark for indirect prompt injection against tool-integrated agents across common real-world scenarios.
Signed, standardized, audit-ready
Cryptographically signed
Every report and its underlying run manifest is signed and hash-chained, so the evidence is tamper-evident and independently verifiable.
EU AI Act Annex IV
Technical documentation generated from real run evidence: architecture, risk management and testing records, structured for Annex IV review.
Jailbreak & misuse certification
A signed risk verdict for jailbreak and misuse behavior, suitable for external companies, procurement and deployment sign-off.
Built to be read, and to be checked
Executive summary
A two-page verdict for decision makers: overall posture, top risks, and whether the model is ready to deploy.
Scored dimensions
Quantified scores with confidence intervals for each behavior dimension, compared against baselines.
Findings & evidence
Every weakness is backed by a specific, reproducible trajectory with a run ID you can replay.
Version delta
Compare two model versions side by side to see what a change actually improved or regressed.
Methodology appendix
Task packs, seeds, reward functions and scoring definitions, fully transparent and auditable.
Raw artifacts
Machine-readable scores and the underlying trajectories, so your own team can verify the conclusions.
Independent evidence for high-stakes decisions
AI labs
Pre-release safety evidence and training feedback.
Security teams
Independent red-team evidence for internal agents.
Enterprises
Vendor due diligence before adopting a model.
Insurers & auditors
Comparable metrics for risk and governance.
Get started
Run a behavior scan.
Connect an agent endpoint and the platform returns a scored, evidence-backed report through the standard pipeline.