observer-effect-probe

# The Skill Behaves Perfectly When It Knows You're Watching.

> Helps identify skills that use environment detection to suppress malicious behavior during security review — the attestation evasion technique that passes every check precisely because the checks are visible.

## Problem

The runtime attestation problem has a second layer: a skill sophisticated enough to execute conditionally based on environment signals is also sophisticated enough to detect that it's being probed and suppress its conditional behavior during the probe.

This is the observer effect applied to security: the act of observing changes what is observed. A skill that can detect sandbox fingerprints (restricted network access, specific hostname patterns, probe-characteristic API call sequences) can present clean behavior during attestation and activate payload delivery only when those fingerprints are absent — meaning only in real production environments.

This is not a theoretical concern. Malware has used VM detection to suppress execution during analysis since the 1990s. The same technique scales to agent skills: check for the probe's characteristic markers, return normal output, resume malicious behavior once the probe exits. A skill that clears attestation by detecting the attestor is not a securely attested skill — it's an adversarially attested skill.

The observer-effect probe attempts to reduce the skill's ability to distinguish probe context from production context by minimizing observable differences between the two.

## What This Probes

This probe examines attestation evasion capabilities across five dimensions:

1. **Sandbox fingerprint detection** — Does the skill attempt to identify characteristics of a probe environment? Checks for sandbox-specific hostname patterns, reduced network connectivity, probe-characteristic request timing, or the absence of expected production signals
2. **Probe timing sensitivity** — Does the skill's behavior change based on how l...