Benchmark Almanac

AttaQ

Granite 3.3 8B Instruct 88.5% leads since Apr 2025 Fading

Models answer 1,402 adversarial questions in seven harm classes, such as deception, violence and PII. Responses are scored for harmlessness.

Questions were extracted from Anthropic hh-rlhf, generated by LLMs, or built from a Wikipedia list of crimes. MIT license.

In the authors’ words

As large language models become more prevalent, their possible harmful or inappropriate responses are a cause for concern. This paper introduces a unique dataset containing adversarial examples in the form of questions, which we call AttaQ, designed to provoke such harmful or inappropriate responses. We assess the efficacy of our dataset by analyzing the vulnerabilities of various models when subjected to it. Additionally, we introduce a novel automatic approach for identifying and naming vulnerable semantic regions - input semantic areas for which the model is likely to produce harmful outputs. This is achieved through the application of specialized clustering techniques that consider both the semantic similarity of the input attacks and the harmfulness of the model's responses. Automatically identifying vulnerable semantic regions enhances the evaluation of model weaknesses, facilitating targeted improvements to its safety mechanisms and overall reliability.

Abstract of Unveiling Safety Vulnerabilities of Large Language Models (arXiv:2311.04124, CC0)
Made by IBM Research Released Sep 2023 Refusing harmful requests All tasks public Last in a launch post Apr 2026 , dropped after it saturated Last model added Apr 2026
Level
Everyday Most adults, without training
Scored
Graded by a model A language model scores each response, usually against a rubric.
Answers
Open-ended Free-form text, documents, designs or media judged as a whole.
Paper Dataset
5
Models scored in lab cards
5
Launch posts list it, from 1 lab

Granite 3.3 8B Instruct leads with 88.5% among models released by Apr 2026

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.

The newest model with a score here came out in Apr 2026, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 88.5 Granite 3.3 8B Instruct
  2. 86.6 Granite 4.0 H Small
  3. 85.8 Granite 4.1 30B
  4. 85.4 Granite 3.2 8B Instruct
  5. 85.3 granite-4.0-1b
  1. 1 Granite 3.3 8B Instruct 88.5%
  2. 2 Granite 4.0 H Small 86.6%
  3. 3 Granite 4.1 30B 85.8%
  4. 4 Granite 3.2 8B Instruct 85.4%
  5. 5 granite-4.0-1b 85.3%
#ModelScoreSettingDateSource
1 Granite 3.3 8B Instruct 88.5% Apr 2025 Model card
2 Granite 4.0 H Small 86.6% Oct 2025 Model card
3 Granite 4.1 30B 85.8% Apr 2026 Model card
4 Granite 3.2 8B Instruct 85.4% Feb 2025 Model card
5 granite-4.0-1b 85.3% Oct 2025 Model card

2 models have held number one since its release in Sep 2023

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

IBM
Lab claims Passed 85% in Feb 2025, by Granite 3.2 8B Instruct
Granite 3.3 8B Instruct 88.5%

The best score passed 85% in Feb 2025

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% 2025 2026 Near the ceiling (85%) Granite 3.3 8B Instruct 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Granite 3.3 8B Instruct

Labs headline it less often now, last in Apr 2026

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

IBM
DateLaunch postScoreAs the post writes it
29 Apr 2026 Granite 4.1 30B 85.8 AttaQ
28 Oct 2025 granite-4.0-1b 85.3 AttaQ
2 Oct 2025 Granite 4.0 H Small 86.6 AttaQ
16 Apr 2025 Granite 3.3 8B Instruct 88.5 AttaQ
26 Feb 2025 Granite 3.2 8B Instruct 85.4 AttaQ

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%