Watcher monitors your agent sessions to detect:
Apollo Research has conducted pre-deployment model evaluations and AI governance work for leading labs, governments, and foundations.
Connect your agents. Watcher surfaces the risk.
Watcher plugs into how you use agents — model routers, proxy layers, or local deployments. It ingests agent sessions, tool calls, and telemetry, attaches user and agent identity, and analyzes behavior in real time to surface the risks that matter.
What we're working on
We're building integrations so Watcher feeds directly into the tools your security team already trusts — SIEMs like Microsoft Sentinel, Splunk, or Elastic, plus alerting and case-management workflows. If there's a specific tool or workflow you need Watcher to plug into, we'd love to hear it.
From detection at scale to real‑time intervention
Surface the failures hiding across sessions
Built for security teams. Monitor agent behavior across your org, surface failure patterns, and define the guardrails your engineers work within — with live alerts when something needs attention.
See the full picture across your org. Watcher flags the highest-priority failures across all agent sessions — by category, frequency, and trend.
Get alerted where your team already works. Critical findings go straight to Slack or Teams so security teams can triage without switching tools.
Trace root cause in one click. Drill into any flagged session to see the full trajectory — user, agent, project — and remediate immediately.
Your security team defines the guardrails.
Your engineers keep moving fast.
Real-time oversight without the friction
Built for engineers. A safer alternative to --dangerously-skip-permissions — monitor active sessions, get alerts on risky trajectories, and intervene before something goes wrong.
Track and manage multiple agents at once. See all active coding agent sessions in one place — what they're doing, where they're blocked, and where your input is needed. Prioritize attention across sessions instead of context-switching blindly.
Step in only when it matters. Watcher surfaces the ~3% of actions that actually require your attention — like accessing secrets or running risky commands. Approve, deny, or guide the agent in real time without slowing everything else down.
Four categories of risk, one unified view
Insecure code execution
- Remote code execution and destructive operations
- Introducing vulnerabilities or exposing secrets
- Unauthorized modifications to production systems
Data exfiltration & access abuse
- Leaking PII, credentials, or internal data to external services
- Privilege escalation and unauthorized access
- Pulling sensitive data across tool boundaries
Agent manipulation
- Prompt injection and goal hijacking
- Memory and context poisoning across sessions
- Circumventing safety constraints or instructions
Emergent & unknown risks
- Cascading failures across agent workflows
- Novel failure patterns that don't trigger rules
- Anomalies only visible at scale across sessions
Guided deployment available now
We'll walk you through setup, data flow, and the right monitoring configuration for your environment.
Book a call