Skip to content

AI Agent Security Audit

Automated AI agent security audit across your working tree and full git history: secret management, API key exposure, prompt injection surface.

Three Critical Attack Surfaces

Each audit checks every surface your AI agent relies on — secrets, keys, and prompts.

Secret Management

We audit how your agent stores, retrieves, and rotates credentials — environment variables, encrypted vaults, hardcoded secrets.

  • Hardcoded credentials detection
  • .env file exposure risk
  • Encryption strength audit
  • Secret rotation policy

API Key Exposure

We scan for leaked, exposed, or improperly scoped API keys — from code repositories, logs, and runtime environments.

  • Git history key leak scan
  • Key scope & permission audit
  • Runtime key exposure checks
  • Third-party key risk assessment

Prompt Injection Surface

We test your agent's vulnerability to prompt injection — direct and indirect — and measure resistance to jailbreak attempts.

  • Direct injection resistance test
  • Indirect injection surface map
  • Jailbreak resilience scoring
  • Output sanitization audit

The AI Security Landscape

Three recent events make AI agent security the most urgent audit for any deployment in 2026.

GLM 5.2 Benchmark

72% Failure Rate

GLM 5.2 security benchmark showed 72% of AI agents fail basic secret management checks. The benchmark tested 200+ production agents — only 56 passed.

Brown University Study

AI-Generated Fraud

Brown University researchers demonstrated AI agents generating fraudulent transactions in 87% of test scenarios when prompted with indirect injection techniques.

ChatGPT Legal Precedent

Legal Admissibility Crisis

ChatGPT outputs were presented as legal evidence in a U.S. court — raising questions about the integrity and provenance of agent-generated outputs when security is unverified.

How an Audit Is Produced

We scan your code. Every finding below carries a file, a line and the matched text, so you can check our work instead of taking it on trust.

01

You send the repository

A public repository URL, or a private one you grant read access to for the duration of the scan. Tell us the branch and what to exclude. Nothing else is needed — no setup, no agent configuration, no integration.

02

We scan the code and its history

Pattern rules run over the working tree: credential exposure, keys in configuration and logs, prompt-injection sinks and delimiter injection. From $75 the git history is scanned too, which is where a rotated key usually survives.

03

You get an HTML report

Findings by severity, each with file, line, the matched text and a remediation note. The report also lists what was excluded and why, and what the scan did not cover — a report that hides what it skipped cannot be audited. Save it as PDF from the browser, or ask for a PDF and we will convert it.

04

You fix, we re-scan at no extra cost

Tell us what the report missed or got wrong. Every re-scan is included in the price until the result meets the agreed quality — fixing findings and verifying the fixes are part of the work, not a second sale. From $250 the report also adds a suppression audit: every match we discounted, with the category that discounted it.

Three Published Tiers

Every tier checks all three attack surfaces. The prices are fixed — the difference is coverage, not formatting.

Basic
$29

Working tree scan — credentials, keys, injection sinks

  • Secret Management
  • API Key Exposure
  • Prompt Injection Surface
  • You get an HTML report
Full
$75

Adds git history — rotated and deleted secrets

  • Secret Management
  • API Key Exposure
  • Prompt Injection Surface
  • Git history key leak scan
  • You get an HTML report
Hardened
$250

Adds a suppression audit — every discounted match shown and justified

  • Secret Management
  • API Key Exposure
  • Prompt Injection Surface
  • Git history key leak scan
  • Every match we discounted, with the reason
  • You get an HTML report

The samples come from a synthetic repository we wrote for this page: every credential in it is fabricated and worthless, so nothing here is a real finding about us or about anyone else.

We deliver first, you pay after. We run the scan and send you the report; when it meets the agreed quality we email a bank-transfer invoice. Re-scans until the result is right are included in the price. This is not a card checkout, and there is no automatic download.

Common Questions

What do I need to provide?

Read access to your repository for the duration of the scan, plus a note on the branch and any paths to exclude. From $75 the git history is scanned as well. We do not need your production systems, and we do not keep the code once the report is delivered.

Do you access my code?

Yes, for the duration of the scan. The scanner reads your files and, from $75, your commit history — that is how a finding can carry a file and a line you can check. We never touch your production systems, and we do not keep the code after the report is delivered.

What format is the report?

HTML report, print-ready: executive summary, counts by severity, per-finding evidence and remediation, plus the excluded matches and the coverage limits. Save it as PDF from the browser, or ask us to convert it.

How long does it take?

The scanner is local, so a scan takes minutes rather than days. We send you the report as soon as it is produced, and confirm the delivery time when you write to us.

Can I order a follow-up scan?

Yes, and it is already inside the price. Re-scans until the result meets the agreed quality are included in every tier — you do not pay a second time for the same audit.