Benchmark Almanac

Prompt injection robustness against adaptive attackers, coding (Anthropic)

Claude Opus 4.6 0% leads since Feb 2026 Only in system cards

Coding agents face indirect prompt injections crafted by an adaptive attacker. Score is attack success rate.

Anthropic uses Gray Swan's Shade adaptive attacker for this surface. Reported at 1 and many attempts.

Made by Anthropic Released Nov 2025 Jailbreaks and prompt injection Tasks kept private Last in a system card Oct 2026 Last model added Oct 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
13
Models scored in lab cards
13
System cards report it, from 1 lab

3 models share the top score of 0%

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

  1. 1 Claude Opus 4.6 0%
  2. 2 Claude Sonnet 4.6 0%
  3. 3 Claude Mythos Preview 0%
  4. 4 Claude Haiku 5.5 0.08%
  5. 5 Claude Opus 4.5 0.30%
  6. 6 Claude Sonnet 5 0.31%
  7. 7 Claude Fable 5 0.45%
  8. 8 Claude Opus 5 0.56%
  9. 9 Claude Opus 4.7 2.34%
  10. 10 Claude Sonnet 5.5 3.01%
  11. 11 Claude Opus 4.8 7.03%
  12. 12 Claude Fable 5.1 9.26%
  13. 13 Claude Opus 5.5 54.6%
#ModelScoreSettingDateSource
1 Claude Opus 4.6 0% Feb 2026 Model card
2 Claude Sonnet 4.6 0% Feb 2026 Model card
3 Claude Mythos Preview 0% Apr 2026 Model card
4 Claude Haiku 5.5 0.08% Oct 2026 Model card
5 Claude Opus 4.5 0.30% Nov 2025 Model card
6 Claude Sonnet 5 0.31% Jun 2026 Model card
7 Claude Fable 5 0.45% Jun 2026 Model card
8 Claude Opus 5 0.56% Jul 2026 Model card
9 Claude Opus 4.7 2.34% Apr 2026 Model card
10 Claude Sonnet 5.5 3.01% Sep 2026 Model card
11 Claude Opus 4.8 7.03% May 2026 Model card
12 Claude Fable 5.1 9.26% Sep 2026 Model card
13 Claude Opus 5.5 54.6% Sep 2026 Model card

2 models have held number one since its release in Nov 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Lab claims 2 leaders
Claude Opus 4.6 0%

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 Claude Haiku 5.5 Claude Opus 4.5 0% 20% 40% 60% 80% 100% 2026 Claude Haiku 5.5 Claude Opus 4.5

Labs report it only in system cards

We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.

13 system cards report it, from 1 lab

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
7 Oct 2026 Claude Haiku 5.5 system card 0.08 Shade adaptive attacker, coding, attempt-level attack success rate without safeguards, 6/40 scenarios. With probes 0% (Table 5.2.2.1.A) Robustness
28 Sep 2026 Claude Sonnet 5.5 system card 3.01 Shade adaptive attacker in coding, attempt-level attack success rate without safeguards, with thinking, 16/40 scenarios (2.63% with probes) (Table 5.2.2.1.A) Robustness
22 Sep 2026 Claude Opus 5.5 system card 54.6 Shade adaptive attacker in coding, attempt-level attack success rate without safeguards, with thinking, 37/40 scenarios (11.13% and 36/40 with probes), lower is better (Table 5.2.2.1.A) Robustness
1 Sep 2026 Claude Fable 5.1 system card 9.26 Shade indirect prompt injection in coding, attempt-level attack success rate, with thinking, without PI probes, including cyber classifier fallback to Opus 4.8 (no attacks succeeded on the 369 requests Fable 5.1 answered directly); 31/40 scenarios; 2.05% with PI probes; new stronger Shade attacker 56.87% (12.80% with probes) (Tables 5.2.2.1.A and 5.2.2.1.B) Robustness
24 Jul 2026 Claude Opus 5 system card 0.56 Shade adaptive attacker in coding, attempt-level attack success rate without safeguards, with thinking, 13/40 scenarios (0.18% and 4/40 with probes), lower is better (Table 5.2.2.1.A) Robustness
30 Jun 2026 Claude Sonnet 5 system card 0.31 Shade adaptive attacker in coding, attempt-level attack success rate without safeguards, with thinking, 7/40 scenarios (0.29% without thinking. 0.09% with safeguards), lower is better (Table 5.2.2.2.A) Robustness
9 Jun 2026 Claude Fable 5 system card 0.45 Shade indirect prompt injection in coding, attempt-level attack success rate without safeguards, with thinking, 40 scenarios x200 attempts, row labeled Mythos 5, 8/40 scenarios broken; with safeguards 0.41% and 11/40, lower is better (Table 5.2.2.1.A) Robustness
28 May 2026 Claude Opus 4.8 system card 7.03 Shade adaptive attacker in coding, 40 scenarios, single-attempt attack success rate without safeguards, with thinking (200 attempts 57.5%. With safeguards 2.09% and 37.5%. Without thinking 17.44%), lower is better (Table 5.2.2.2.A) Robustness
16 Apr 2026 Claude Opus 4.7 system card 2.34 Shade attacks in coding, with thinking, without safeguards, 1 attempt. 200 attempts 60.0. Without thinking 10.43 and 92.5. With safeguards 0.43 and 25.0 (Table 5.2.2.1.A) Robustness
7 Apr 2026 Claude Mythos Preview system card 0 Shade attacks in coding, extended thinking, without safeguards, 1 attempt, lower is better; 0.0 at 200 attempts, standard thinking 0.03 and 2.5 (Table 8.3.2.2.1.A) Robustness
17 Feb 2026 Claude Sonnet 4.6 system card 0 Shade attacks in coding, extended thinking, without safeguards, 1 attempt; 0.0 at 200 attempts, standard thinking 0.1 and 7.5 (Table 5.2.2.1.A) Robustness
5 Feb 2026 Claude Opus 4.6 system card 0 Shade adaptive prompt injection in coding, attack success rate, extended thinking, without safeguards, 1 attempt, lower is better; 0.0 in every condition including 200 attempts (Table 5.2.2.1.A) Robustness
24 Nov 2025 Claude Opus 4.5 system card 0.3 Shade adaptive indirect prompt injection, coding, attack success rate at 1 attempt, extended thinking, no safeguards, lower is better (Table 5.2.2.1.A; 10.0% with 200 attempts, standard thinking 0.7% and 17.5%) Robustness

More in Safety and alignment

All 444 in Safety and alignment
  • FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
  • StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is. Inkling 98.6%
  • AgentDojo In use Agents do tool-use tasks while tools and data carry prompt injections. Attack success rate counts hijacked runs. Grok 4 Fast 0%
  • MASK In use Measures whether a model contradicts its own elicited beliefs under pressure to lie. Honesty separated from accuracy. Muse Spark Open 98.1%
  • Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. GPT-6 Astra 8.50%
  • Agentic misalignment suite (Anthropic) In use Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate. Claude Sonnet 4.5 0%