Ship AI agents you can trustverifydefendproveship

Aegvale red-teams your LLM agents for jailbreaks, prompt injection, tool abuse and data leakage - then hands you prioritized, reproducible findings and a security score you can gate releases on.

329
adversarial tests
31
attack suites
OWASP
LLM Top 10 mapped
MITRE
ATLAS techniques

Mapped to the standards your auditors and customers ask about

OWASP LLM Top 10MITRE ATLASNIST AI RMFGDPRSOC 2-alignedISO 27001-alignedISO 42001

Inside the product

The full report, the moment your scan finishes

Security score, severity breakdown and reproducible findings - everything waiting for you the moment you connect an agent.

aegvale · support-agent - scan complete
68/ 100
Not production-ready +12 vs last scan
329
Tests run
20
Findings
3
Critical
71%
Pass rate

Findings by severity

Critical
3
High
5
Medium
8
Low
4

Top findings

mapped to OWASP LLM Top 10
System-prompt extraction via role-playLLM01
Tool call: transfer_funds with attacker argsLLM06
Indirect injection through retrieved documentLLM01
Adversarial coverage

Every jailbreak, injection and tool-abuse technique - run on demand.

scanning support-agent…
PASSRefusal held · jailbreak / DAN-12JB
FAILSystem prompt leaked · role-playLLM01
FAILTool call: transfer_fundsLLM06
PASSRedacted PII on requestLLM02
FAILObeyed injected retrieved docLLM01
Ship-blocking gate

Fail the build when your agent regresses - right in CI.

Aegvale security scan

Required · score 68 < 80 · 3 critical

Blocked

Unit tests

142 passed

Passed

Coverage across every surface an attacker can reach.

Prompt
Tools
RAG

Ship-blocking, not advisory

Gate every release on a security score

Drop Aegvale into CI and fail the build when your agent regresses. One command, and no risky release slips through.

  • Runs on every pull request
  • Fails on new critical/high findings
  • Posts a summary back to the PR
.github/workflows/security.yml
- name: Aegvale security scan
  uses: aegvale/scan-action@v1
  with:
    api_key: ${{ secrets.AEGVALE_API_KEY }}
    project: support-agent
    fail_on: high      # block the build
    min_score: 80

Workflow

From endpoint to evidence in three steps

01

Connect your agent

Point Aegvale at any HTTP agent or model endpoint. Credentials are stored encrypted at rest.

02

Run a scan

Choose quick or deep coverage across OWASP LLM Top 10 and MITRE ATLAS techniques.

03

Ship with confidence

Get prioritized findings with CVSS, reproduction bundles and remediation SLAs you can act on.

Security & compliance

The platform that tests your security holds itself to the same standard

Encryption, least privilege and verifiable audit trails are built in - so security and compliance teams can sign off with confidence.

Encrypted secrets

Agent API keys are encrypted at rest with AES-256-GCM.

Strong authentication

Two-factor auth, SSO and rotating refresh tokens.

Least privilege

Role-based access across Owner, Admin, Member and Viewer.

Auditable by design

Signed, independently verifiable audit exports for every scan.

FAQ

Questions, answered

What do I need to connect?

Any AI agent or model behind an HTTP endpoint: OpenAI, Anthropic, Google, Azure, Mistral, Ollama, or a custom REST API. You bring your own model keys; they are encrypted at rest.

How is this different from a normal pentest?

It is automated, repeatable and built specifically for LLM agents (jailbreaks, prompt injection, tool abuse and RAG poisoning), so you can run it on every change instead of once a year.

Will it leak my data?

Scans run against your own endpoint with your own keys. Secrets are encrypted (AES-256-GCM), access is role-based, and every scan produces a signed, verifiable audit trail. A self-hosted option is available.

Can I gate deployments on it?

Yes. Run Aegvale from the CLI or a GitHub Action and fail the build when the security score drops below your threshold.

Background Paths
Background Paths

Findtheexploitfirst.

Start testing your agents in minutes. No agent is too small to red-team.

Start free