Summary

With 99.8% of prompts carrying technique classifications (140,977 of 141,223), this report presents the first comprehensive technique-level ASR analysis. 104 techniques have sufficient graded results (n >= 10) and 155 technique-by-model-family pairs (n >= 5).

Corpus snapshot: 50,182 graded results across 104 techniques, tested on models from 10+ families.

Top 10 Most Effective Techniques (Broad ASR)

Rank	Technique	Era	n	Broad ASR
1	compliance_cascade	reasoning_2025	20	85.0%
2	obliteratus_probe	general	42,346	81.2%
3	temporal_drift_attack	reasoning_2025	39	74.4%
4	format_lock	reasoning_2025	130	70.8%
5	infrastructure_boundary	reasoning_2025	10	70.0%
6	cot_manipulation	reasoning_2025	22	68.2%
7	hcot_attack	reasoning_2025	29	62.1%
8	physical_harm	general	17	58.8%
9	urgency_manipulation	general	18	55.6%
10	faithfulness_gap_exploit	reasoning_2025	443	51.0%

7 of the top 10 techniques are from the reasoning_2025 era, suggesting reasoning-model-specific attacks are disproportionately effective. format_lock (70.8%, n=130) and faithfulness_gap_exploit (51.0%, n=443) are the most robust findings by sample size.

---

Report #313 | F41LUR3-F1R57 Adversarial AI Research