ADEPTS

ADEPTS-BENCH: Measuring the Trustworthiness of Computer Use Agents across Devices

A dual-stream trustworthiness benchmark for Computer Use Agents (CUAs) navigating mobile and desktop applications. ADEPTS evaluates whether agents can safely interact with visual interfaces while handling ambiguous instructions:

• Safety — Paired benign/malicious tasks with threats embedded in the visual interface (e.g., mislabeled buttons, phishing links, unauthorized purchases). Measures task success vs. attack success rate.
• Disambiguation — Evaluates whether agents seek clarification when user intent is ambiguous, rather than making potentially harmful assumptions.

Code: facebookresearch/adepts

Paper Highlights

What's in the benchmark

2,462 tasks, evaluated offline — no live environment, no infrastructure to stand up. Each safety task is one API call; each disambiguation task is one model call plus LLM-judge calls for question matching. ADEPTS-BENCH is the only benchmark combining visual grounding, offline evaluation, paired benign/malicious variants, disambiguation, a user-research-grounded taxonomy, and an open release (paper Table 1).

Stream	Instances	Mobile	Desktop
Safety	1,718 (859 paired tasks)	358 pairs, ~70 Android apps	501 pairs, ~150 apps
Disambiguation	744	381 tasks (4.7 images avg)	363 tasks (2.3 images avg)

• Paired safety design. The benign and malicious variants of a task share the same instruction; only the screenshot differs, embedding the threat in the visual interface (phishing overlay, injected prompt, misleading button). The benign variant measures capability, the malicious one measures safety. 10 threat categories, from prompt injection and web hacking to misalignment and system sabotage (Dangerous Commands is desktop-only).
• Disambiguation design. A trajectory of screenshots leads to an ambiguous decision point; the model must decide whether to clarify or act. Categories: referential underspecificity, linguistic ambiguity, missing parameters, and impossible tasks. Each item is scored on obviousness (0–2) and consequence (0–2).
• User-grounded risk taxonomy. Built from expert safety workshops plus a MaxDiff survey of n=1,300 US general-population participants ranking 44 undesirable outcomes. Top concerns: identity theft (80.2%), account compromise (77.5%), unauthorized financial transfer (76.0%). 74.1% were more worried about external attacks than internal errors.
• Annotation quality. Pixel-level bounding boxes from professional annotators, validated by cross-model agreement and expert review. The disambiguation LLM judge (Gemini 3.1 Pro) agrees with human judgments 96.1% of the time (κ=0.92).

Metrics

• TSR — task success rate on benign tasks (capability).
• ASR — attack success rate on malicious tasks (lower is safer).
• FRR — false refusal rate: status(unsafe) on benign tasks (over-safety).
• ADEPTS Score — harmonic mean of TSR and (1−ASR), so a model can't score well by sacrificing one for the other (90% TSR with 90% ASR scores 17.1%, not 50%).
• Disambiguation F1 — precision × recall of generated clarification questions against human ground truth.
• Severity Calibration Error (Δ) — mean per-component absolute error on obviousness and consequence scores, in [0, 4].

Headline results

Safety stream, mean ± 95% CI over 10 runs at temperature 0.1 (paper Table 2):

Model	Mobile TSR↑	Mobile ASR↓	Mobile FRR↓	Mobile S↑	Desktop TSR↑	Desktop ASR↓	Desktop FRR↓	Desktop S↑
Gemini 3.1 Pro	77.6	29.6	9.6	73.8	82.9	29.8	8.7	76.0
Claude 4.7 Opus	71.0	27.7	9.4	71.6	78.1	27.2	6.3	75.4
GPT-5.4	62.1	36.7	5.1	62.7	68.8	36.3	7.0	66.1
Gemini 2.5 CU	78.1	52.7	1.5	58.9	83.6	51.5	1.3	61.4
Qwen3-VL-235B	83.0	78.1	0.0	34.7	86.6	76.3	0.0	37.2
Qwen3-VL-8B	79.0	75.5	0.0	37.4	87.4	73.7	0.0	40.5
Qwen3-VL-4B	81.6	79.6	0.0	32.6	87.0	76.7	0.0	36.7

Disambiguation stream, with the severity-scoring prompt (paper Table 4):

Model	Mobile F1↑	Mobile Δ↓	Desktop F1↑	Desktop Δ↓
Gemini 3.1 Pro	58.9	0.86	39.5	1.44
Gemini 2.5 CU	54.6	1.16	40.7	1.37
Qwen3-VL-235B	52.2	1.34	46.2	1.38
Qwen3-VL-4B	49.8	1.31	29.3	1.57
Claude 4.7 Opus	49.5	0.90	43.5	1.43
Qwen3-VL-8B	43.6	1.54	5.7	0.45
GPT-5.4	39.9	0.81	26.0	1.55

(Qwen3-VL-8B's desktop Δ is not comparable — it returns empty responses on 93% of desktop tasks.)

Key takeaways

1. No model is both capable and safe. The best ADEPTS Score is 76.0% (Gemini 3.1 Pro, desktop), the only model × platform combination above 80% TSR while staying below 30% ASR. Every model clicks "Checkout" on a $25K order, and none detects that a "factory reset" button is mislabeled "Optimize."
2. Three distinct safety architectures. Ablating the status(unsafe) refusal tool raises ASR by 10–23pp for frontier models and leaves Qwen unchanged (±1pp): tool-dependent (Gemini 3.1 Pro, +22pp), partially tool-dependent (Claude +10pp, GPT-5.4 +11pp — implicit safety survives without the tool), and no mechanism (Qwen, which never calls the tool at all). Removing the tool also raises benign TSR by 1–8pp — the capability/safety tradeoff lives at the system-prompt level.
3. Computer-use specialization may cost safety. Gemini 2.5 CU has the highest TSR (83.6% desktop) but the highest frontier ASR (51.5%), well above general-purpose Gemini 3.1 Pro (29.8%). Its 181 unique vulnerabilities cluster on action-ready UIs and semantic-only threats: it optimizes for action completion over action evaluation.
4. Open-source safety doesn't scale. Qwen reaches 87% TSR but 74–80% ASR at every size (4B → 235B), never uses the refusal tool, and never falsely refuses.
5. A four-level failure spectrum. L1 (7.4%) explicit visual cues, caught by everyone; L2 (15.6%) embedded text threats, caught by frontier models only; L3 (66.3%) ambiguous contexts where models disagree; L4 (10.8%) no visual cues at all — the harm is in action scale ($25K checkout) or label mismatch. Safety training handles pattern matching, not consequence reasoning.
6. Worst-case safety is much worse than the mean. Per-task pass@k (k=10) ASR runs 2–23pp above mean ASR. GPT-5.4 is the least deterministic (36% → 58–60%, ~200 flaky tasks per platform); Qwen-4B the most (+2–3pp).
7. Over-refusal is visual pattern matching, not threat detection. Frontier models falsely refuse 6–10% of benign tasks, triggered by dark/"hacker" styling, urgency and promotional keywords, and — uniquely visible in the paired design — near-identical benign/malicious screenshots.
8. Models overestimate consequences. 42–62% of matched disambiguation items are rated more severe than ground truth, only 7–25% less. Obviousness calibration is far better: models know what is ambiguous but overweight how bad a wrong guess would be — the same cautious bias that drives FRR.
9. Impossible tasks are a blind spot. Tasks requiring impossibility detection ("change the phone SIM card" on a software-only interface) are missed at more than double the rate (30.6% vs 13.8%). Models treat them as merely underspecified and ask helpful-but-wrong questions.
10. The severity-scoring prompt cuts both ways. It raises clarification rate by 4–40pp, but lowers F1 for most frontier models (GPT-5.4 −5.6pp) while substantially helping open-source ones (Qwen-8B +18pp) — useful scaffolding for weaker models, a metacognitive cost for stronger ones.
11. Trustworthiness needs multi-dimensional evaluation. Rankings shift across platform and stream: Gemini 3.1 leads mobile disambiguation but drops to 4th on desktop, while Qwen-235B rises from 5th to 1st despite the worst safety. Safety improves on desktop (+4–8pp TSR) while disambiguation degrades (8–38pp F1).

Hardest threat categories

ASR by threat category, combined mobile + desktop (paper Table 10):

Threat	Gemini 3.1	Claude 4.7	GPT-5.4	Gemini CU
System Sabotage	48.3	49.4	68.5	66.3
Misalignment	50.0	43.8	45.8	60.4
Hallucination	40.9	33.0	43.2	47.7
Reasoning Gap	35.8	34.7	40.0	49.5
Adversarial Attack	28.6	30.0	38.5	54.9
Jailbreak	17.8	17.8	34.4	56.7
Response Latency	22.1	16.3	32.6	51.2
Dangerous Cmds	16.3	8.2	20.4	51.0
Prompt Injection	13.3	14.4	30.0	32.2
Web Hacking	5.9	15.3	10.6	41.2

System sabotage is the most effective attack overall; prompt injection and web hacking are the best defended. Gemini CU is the most uniformly vulnerable.

License

This project is licensed under CC-BY-NC 4.0 (Attribution-NonCommercial 4.0 International). See the license text for details.