Describe the agent
Tell us what it accesses, what it can do, and how independently it operates.
Evidence-based AI-agent security assurance
Know what your AI agent can break before it does.
Combine a 25-control risk assessment, a customer-authorised local inspector and controlled adversarial testing. Receive declared-risk analysis, technical evidence, repeated adversarial trials, reproduced failure cases, credible attack paths and a decision-ready remediation report.
32 controlled attack cases1-5 repeated trials per caseNo source, secrets or raw transcripts uploaded
PERMISSIONSPROMPT INJECTIONMCP TOOLSSECRETSAUTONOMYMEMORY POISONINGTOOL ABUSEINCIDENT RESPONSE
From declared controls to observed evidence.
The system keeps self-declared risk separate from locally observed technical evidence. It does not label a static scan as a penetration test or certification.
Measure risk and confidence
Separate inherent exposure, control weakness and the strength of the evidence behind every answer.
Inspect the implementation
Run a transparent local scanner for secrets, dependencies, MCP tools, CI/CD, containers and agent-control signals.
Attack the staging agent safely
Run 32 prompt-injection, data-leakage, tool-abuse, memory, authorisation, output and resource-control cases with repeat trials through a dry-run adapter.
Connect attack paths
Combine authority, exposure, static observations and reproduced behaviour into realistic failure scenarios.
Remediate and retest
Use exact controls, evidence requirements, test methods and repeat runs to prove progress.
Evidence ladder
Know exactly what the result proves.
AgentRiskLayer does not blend claims, observations and reproduced failures into one vague score. Each evidence class is labelled and kept traceable.
Declared
The customer describes exposure and controls. Evidence confidence shows whether each claim is unsupported, documented or tested.
Questionnaire evidenceObserved
The read-only Inspector identifies repository, CI/CD, container, MCP, dependency and control signals without uploading source or secret values.
Signed static evidenceReproduced
The customer-operated runner executes controlled attacks against an authorised staging adapter using synthetic data and dry-run tools.
Repeated adversarial evidenceRetested
Changes are compared across runs so teams can prove which findings were resolved, newly introduced or remain open.
Change and closure evidencePurpose-built checks.
Focused landing pages make the assessment discoverable for teams searching for a specific AI-agent security problem.
Founding beta pricing
Start free. Pay only for the evidence you need.
Introductory pricing supports the controlled beta while we measure false positives, remediation success and customer outcomes. Prices may change for future customers.
Free score
£0
A fast first-pass security picture.
- 0–100 risk score
- Top three declared findings
- Local inspector download
- Private by default
Essential report
£9.99
Complete findings and remediation.
- All declared findings
- Technical inspection summary
- Prioritised recommendations
- 30-day action plan
- PDF and email delivery
Professional report
£24.99
Decision-ready evidence for a launch review.
- Everything in Essential
- Complete technical evidence register
- 2 authorised red-team campaigns
- 32 cases with repeat trials
- Reproduced failure evidence
- Deployment decision and retest criteria
Developer
£19/month
Repeat professional assessments.
- Professional reports
- 10 controlled runs / 30 days
- Saved assessment history
- Subscription management
- Renewal invoices via Stripe
Clear boundaries. No security theatre.
The score is designed to help teams identify where deeper review is needed—not to create a false certification.
Private by default
Verified accounts, MFA support, private assessment links and explicit sharing controls protect customer evidence.
Explainable scoring
Every risk point maps to the answer supplied, a finding and a concrete control recommendation.
Integrity and lifecycle controls
Bundles are signed, replay-protected, scope-bound and automatically deleted when their approved retention period expires.
Controlled, not reckless
The red-team runner refuses production targets and destructive actions. It uses synthetic data and dry-run tools, and clearly separates a pipeline simulation from evidence about a staging target.