Skip to content

MITRE ATLAS case-study draft — embodied VLA red-team

Draft for review — not submitted. MITRE ATLAS accepts community case studies via GitHub PR to atlas-data (STIX 2.1). This is the draft we'd propose; opening the PR is gated on sign-off.

ATLAS (https://atlas.mitre.org) is the adversarial-ML ATT&CK analog and is actively soliciting agent/embodied techniques — its embodied coverage is thin, which is the opening.

Case study: instruction-jailbreak of a VLA policy in simulation

  • Summary. A templated instruction reframe (RoboPAIR-style) redirects a Vision-Language-Action policy's end-effector out of its benign envelope in simulation, measured by Provael as an Attack Success Rate with a benign-FPR control.
  • Target. Open VLA policy (e.g. SmolVLA) on the LIBERO benchmark.
  • Procedure. provael reproduce robopair --policy smolvla --suite libero (gated, GPU).

ATLAS technique mapping (proposed)

Now structured (D5). This mapping is promoted from prose to the atlas_techniques field on each provael.eai.CATALOG entry and surfaced into SARIF rule properties.atlasTechniques, so external validation routes through MITRE ATLAS with zero Top-10 branding conflict (INV-6). The table below is the human-readable mirror; phrasings are descriptive tactic→technique (not invented AML.TXXXX ids) and flagged (proposed) where ATLAS's thin embodied coverage has no on-point technique.

Provael attack family EAI ATLAS tactic → technique (proposed)
instruction (jailbreak) EAI01 ML Attack Staging → prompt-injection / jailbreak of an ML-driven agent
visual (patch / decoy / sensor-spoof) EAI02 Evasion → adversarial example in the perception channel (craft adversarial data)
backdoor (object / phrase trigger) EAI03 Persistence → backdoor the ML model; ML Supply Chain Compromise → poison an open-weights checkpoint
action / action_space (freeze / hijack) EAI04 Impact → manipulate / deny the agent's actuation (proposed — ATLAS embodied coverage is thin)
injection (scene-text / tool-desc) EAI05 Execution → indirect prompt injection via the environment
misalignment (embodiment gap) EAI06 Impact → unsafe embodied action under a language-benign instruction (proposed — no on-point ATLAS technique)
authorization (excessive autonomy) EAI08 Privilege Escalation → excessive agency / self-authorized guarded action (proposed)
confidentiality (extraction / membership) EAI09 Exfiltration → model extraction / membership inference via the inference interface

Evidence

The Provael run emits an AVID record (provael export --format avid) and OSCAL assessment-results; ATLAS case-study fields (summary, target, procedure, references) are drawn from those plus docs/TOP10.md. RoboPAIR (arXiv:2410.13691) and BadRobot (arXiv:2407.20242) are the canonical robot-jailbreak references to align to.