Skip to content

Risk Detection

Omnodex includes a declarative rule engine that analyzes traced events and flags behavior that may indicate security risks, policy violations, or unexpected agent actions.

Detection rules are declarative definitions - they describe a pattern to match and what it means when matched. The rule engine evaluates each event against all active rules and produces findings.

Rules operate on single events (point-in-time analysis). Each rule specifies:

  • What to match - Which event types and field patterns trigger the rule
  • Risk level - How severe the finding is (critical, high, medium, low)
  • Description - Human-readable explanation of why this matters

The community rule set covers these categories:

Detects when an agent reads files that typically contain secrets or sensitive system information. Examples: SSH keys, environment files, password databases, credential stores.

Detects API keys, tokens, or passwords appearing in tool parameters where they should not be.

Detects credentials being sent to unexpected external destinations. This catches scenarios where an agent might inadvertently (or intentionally, if compromised) send secrets to a third party.

Detects risky package installations, tool shadowing, dependency confusion, hook config manipulation, MCP URL mutations, workspace config access, and IOC domain matching. This is the largest rule category with 9 rules covering the full range of software supply chain attack patterns.

Detects outbound network connections to domains that are not clearly related to the task at hand. Catches supply chain attacks or compromised tools reaching out to command-and-control infrastructure.

Detects cryptocurrency wallet creation activity: CLI tools, wallet APIs, BIP-39 mnemonic generation, private key hex patterns.

Detects untrusted input being passed to sensitive operations without proper validation or sanitization.

Detects patterns where an agent coordinates with or spawns other agent processes in unexpected ways.

Detects when an agent accesses files outside its expected working directory.

Detects agents modifying startup files, cron jobs, scheduled tasks, or other persistence mechanisms.

Detects agents attempting to disable security controls or sandbox restrictions.

Detects agents modifying their own configuration, hook files, or monitoring setup.

Detects execution of known malicious command patterns.

Detects excessive resource usage patterns (runaway loops, disk fills, network floods).

LevelWhen Assigned
criticalActive security threat - credential exfiltration, data breach indicators
highSignificant risk requiring immediate review - exposed credentials, sensitive data access
mediumUnusual behavior worth investigating - unexpected network calls, atypical patterns
lowInformational - logged for audit trail completeness

Community rules (included in the free tier) cover the categories above with general-purpose patterns. They are open source and can be inspected.

Advanced rules (Pro and Enterprise tiers) add machine-learning-based detection, behavioral baselines, and rules that leverage cross-session intelligence. These rules are encrypted and executed server-side.