Apollo Research — the organization that studies how AI models deceive each other at the frontier — shipped a runtime monitor for coding agents three weeks ago. It installs in one command, works with Claude Code and Codex today, and blocks 93% of dangerous tool calls before they execute. That number comes from production data: Apollo Watcher is already monitoring billions of agent tokens per month across enterprise engineering teams.
September Made This Necessary #
September 2026 has been a rough month for AI agent security. GitSpawn dropped on September 1: a single .git/config entry runs attacker code the moment an affected agent runs git status — before any tool-approval prompt appears. Claude Code, Codex, and Cursor are all on the affected list, and four of eight findings remain unpatched. Then on September 9, GreyNoise documented a single attacker using hundreds of AI agents to compromise 440 print servers across 395 organizations in 48 countries — initial compromise of 11 organizations in 26 seconds. Then came the credential harvest: a multi-agent framework that pulled thousands of third-party credentials in under six hours.
The common thread across all three is speed. Agents outpace human review by orders of magnitude. A monitor has to be faster than the agent, not faster than you.
What Watcher Actually Is #
Apollo frames Watcher as “MDM for coding agents,” and the analogy holds. Mobile device management enforces security policies across employee phones — silently, centrally, without interrupting the user — and Watcher does the same for every tool call an agent wants to make. It is not a second chatbot reviewing your prompts. It sits at the hook layer, intercepts before execution, and makes a decision.
The product has two layers. Watcher Live is the real-time blocking layer: a fast general-purpose check runs on every tool call, and anything that trips it escalates to a specialized monitor that can auto-approve, auto-reject with a reason, or page a human. Watcher Analyze is the observability layer: trailing monitors that periodically score the full session transcript across nine risk dimensions and alert when something looks off even if no single action triggered Live.
Apollo built Watcher’s defaults against a library of 40+ failure modes collected from tens of thousands of real-world coding agent transcripts. That is the same empirical research program that produces their frontier model evaluations — the defaults catch what the data says is most likely to go wrong.
What Gets Blocked #
The failure modes Watcher targets read like a list of things that have already happened to production teams this year:
- Credential and secret leakage — agent reads or transmits
.envcontents - Unauthorized file deletion — mass
rmwithout approval - Code pushed without human sign-off — auto-push, force-push
- Scope overreach — agent operates well outside the task it was given
- Deceptive reasoning — stated plan diverges from actual tool calls
- Fabricated test output — made-up pass rates, hallucinated status
These are not edge cases. They are the incidents that have been quietly embarrassing engineering teams since coding agents hit production scale in 2025.
Installing It #
Cloud-hosted setup is one command. See the installation documentation for full details:
curl -fsSL https://github.com/ApolloResearch/watcher-bin/releases/latest/download/install.sh | bash -s -- --cloud
That installs the Watcher client and registers six hooks in ~/.claude/settings.json: session start, user prompt submission, permission requests (blocking), post-tool-use tracking, session end, and graceful stop. The hooks fire automatically on every Claude Code session. For Codex, setup is equivalent. Apollo’s cloud servers are in Western Europe; for organizations with different data residency requirements, self-hosted deployment is available — plan for about an hour of setup and a conversation with Apollo’s team.
Teams already using Tailscale have a faster path. Tailscale’s Aperture integration feeds organization-wide agent logs directly into Watcher, giving engineering leadership visibility across every developer’s agent sessions without per-developer configuration.
The Honest Catch #
Cursor support is not there yet. If Cursor is your primary agent, Watcher is not useful to you today — Apollo says support is in active development. Pricing is not public: the cloud tier appears free to try, but enterprise tiers require a sales conversation, and self-hosted is sales-only. Apollo also has not published a contributor license agreement or code review policy for the GitHub repository yet. None of this is disqualifying for Claude Code and Codex users. Just know what you are walking into.
The Bottom Line #
ByteIota covered GitSpawn and the PaperCut AI agent campaign when they happened. Both showed that the threat surface for coding agents is now being actively exploited at scale. Watcher is the first production-grade, self-serve tool from a credible safety research organization that addresses the problem directly — not by patching individual agent vulnerabilities, but by adding an enforcement layer that persists regardless of which vulnerability gets disclosed next.
If you run Claude Code or Codex in production, the install command is above. One command, thirty seconds, and you have the same monitoring layer that enterprise teams are already running. The coverage in TechCrunch this week was validation, not introduction — Watcher has been in production long enough to have real numbers. September 2026 made the case for it. The question is whether you install it before something on your team does what PaperCut’s attackers’ agents did.