Agent Supervisor

Agents write the code.
Matt secures it.

For mission-critical orgs running AI coding agents in production. One bad diff is one too many. Matt catches the moment an agent drifts. Deletes a failing test. Runs --no-verify. Fakes a green build. Reintroduces a CVE. Every session recorded, every drift named, every run replayable.

seclens research backedrecords · attributes · catchessupervises
claude-code · checkout-servicelive
01 /
why now

Autonomous engineering needs autonomous security.

Human AppSec cannot sit inside every Claude Code, Codex, Cursor, or Windsurf loop. The control has to run where the agent works.

01 / signal

Agents can change production code.

They read tickets, edit files, run tests, open PRs, and increasingly work in the background.

02 / signal

Rewards can be gamed.

Tests passing and tickets closing become optimization targets. Matt verifies the work was earned safely.

03 / signal

Matt catches it inside the loop, not after the PR.

By the time the PR opens, the agent has already deleted the test, neutered the assertion, and faked the green build. Matt sits inside the session, names the drift, kills the run, and produces a replay.

where matt sits

Current AI security companies sit at the edges. Governing what an agent can touch. Gating what tools an agent can call. Matt sits inside the loop. Records what the agent actually did, names how it drifted, replays the moment.

Different telemetry. Different verb. Cross-vendor and independent by design.

02 /
proof

The artifact no other vendor produces.

Every commit, traced back to the exact prompts that produced it. As typed. Attributed to the developer, the session, the machine.

Current AI security companies do not see prompts. Coding agents see only their own. Matt sees them all, attached to commits, as evidence.

commit 7a3f9cmaincheckout-serviceai-authored
feat: add order filtering by name
all 3 prompts from sessionc109d2c4ondevbox-04devsiddharth@mattersec.com
  1. 1tell me the prompt for new session
  2. 2add a filter by name to the order list endpoint
  3. 3i want it pasinated. like 25 per page
attributed190 lines AI · 0 lines human
by modelclaude-opus-4-7 96% · sonnet-4 4%
by toolclaude code 100%
preview from app.mattersec.com · running today inside Appknox
03 /
compatibility

Matt runs alongside the coding agents your team already uses.

Start with the agents active in engineering today. Expand to new background workers and internal agents as your adoption grows.

CursorCursor live
Claude CodeClaude Code live
CodexCodex live
WindsurfWindsurf live
coming soonGitHub Copilot coming soon
coming soonCline coming soon
coming soonKilo Code coming soon
coming soonGemini CLI coming soon
coming soonAntigravity coming soon
coming soonOpencode coming soon
coming soonZed coming soon
coming soonQwen Code coming soon
coming soonInternal agent coming soon
04 /
research foundation

SecLens tells us where models miss vulnerabilities.

MatterSec starts from measurement. SecLens evaluates frontier models on real vulnerability detection tasks and shows that security capability changes by role, language, and vulnerability class.

406CVE-grounded tasks
93open-source projects
10languages
8OWASP categories

SecLens tells us where AI coding models are weak. Matt watches real agent work and catches the moment those weaknesses turn into deleted tests, faked builds, or reintroduced CVEs.

05 /
private beta

Let agents ship faster. Don’t let them drift their way to a green build.

MatterSec is first running inside Appknox. We are opening design partner conversations with teams already adopting AI coding agents in real repositories.