For engineers securing AI-assisted and agent-written code. Agents copy patterns, drop guards to make tests pass, and widen scopes to make deploys green. Aevral reads what they wrote for who has access to what.
AI security engineers watch a new failure shape: code that is fluent, plausible, and quietly wrong about access. Agents copy the pattern next to the one you asked for. They remove the check that made the test fail. They request the scope that makes the deploy green. Each change is small, reads fine, and redefines who has access to what.
The access rules are exactly where agent-written code fails quietly, because a missing gate can survive the tests the agent is optimizing for, which assert behavior, not ownership, tenant, or role boundaries.
Aevral reads the repository snapshot for authorization, IDOR, and business-logic access control, and through its opt-in PR review add-on, it reads every pull request the same way: up to two findings per review, grounded on the added lines, posted as an advisory Check plus inline comments. It never blocks a merge. That shape fits a pipeline where most authors are agents and the human review budget is the bottleneck.
The findings hand back into the same pipeline: each finding carries a fix prompt you can pass to Claude Code, Cursor, or Codex, and a human reviews before merge. The agent that wrote the access bug is a reasonable agent to fix it, once a human has seen the evidence.
The kind of change it reads
A generated route ships with no gate at all
Missing gate
A scope widens to make a deploy pass
Over-grant
A tenant check becomes a prefix match
Business logic
Today Aevral scans authorization, IDOR, and business-logic access control. Not memory corruption, not injection, not a general SAST. A finding is a lead with evidence, and a human decides. PR review is live as an opt-in add-on per organization; paid PR plans are coming soon.
Read next
The PR security review product; For AI agents; Aevral beside AI code reviewers.
Also for
One GitHub App. A report with evidence. A prompt for the agent you already use.