SafePyramid
Models identify rules violated by a conversation under a supplied policy. Rule matching rate averages four thresholds for false positive and false negative counts.
The dataset pairs 1,000 conversations with policies at three levels. Scoring excludes refusals and parsing failures and reports their rate separately.
- Made by
- ByteDance and University of Melbourne
- Released
- June 2026
- Kind
- Refusing harmful requests
- Level
- University
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
- GPT-5.5 #21 on the capability index
- Last model added
- May 2026
- Scores refreshed
- November 18th, 2026
In the authors’ words
In real-world applications, guardrails are often expected to identify unsafe user-model interactions according to application-specific safety policies, rather than relying on predefined risk taxonomies. In this work, we study this setting under the paradigm of in-context policy guardrailing, where guardrails predict safety violations based on policy specifications provided in context. To systematically evaluate this capability, we introduce SafePyramid, a safety benchmark comprising 1,000 multi-turn conversations across 10 domains and 3,000 corresponding application-specific policies, which together contain 61,699 distinct natural-language rules. SafePyramid organizes the evaluation into three difficulty levels: L0 evaluates individual-rule understanding, L1 evaluates reasoning over rule dependencies, and L2 evaluates adaptation of full novel policy frameworks defined in context. To ensure benchmark quality, we employ a rigorous multi-stage pipeline to construct and validate the benchmark. Using SafePyramid, we evaluate 10 frontier LLMs and 5 policy-configurable guardrails and find that in-context policy guardrailing remains highly challenging: even the best-performing model, GPT-5.5, exactly identifies the full set of violated rules in only 54.0%, 35.3%, and 12.9% cases on L0, L1, and L2, respectively. These results highlight the limitations of current guardrails and call for stronger in-context policy guardrails that can reliably execute policies, resolve rule dependencies, and adapt to novel policy frameworks.
GPT-5.5 leads with 54.2% among models released by May 2026 GPT-5.5 leads with 54.2% among models released by May 2026
Best score per entry from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in May 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 GPT-5.5 54.2%
- 2 Claude Opus 4.7 52.9%
- 3 Kimi K2.6 52.3%
- 4
DeepSeek-V4-Pro 52.2%
- 5
Seed2.0 Pro 50%
- 6
Gemini 3.5 Flash 49.6%
- 7 Hunyuan-HY3-Preview 49.1%
- 8
Qwen 3.6 Max (Preview) 47.4%
- 9 GLM-5.1 42.7%
- 10 Grok 4.3 40.6%
- 11 gpt-oss-safeguard-120b 23.6%
- 12 gpt-oss-safeguard-20b 22.5%
Source: Official SafePyramid leaderboard
- 1 GPT-5.5 54.2%
- 2 Claude Opus 4.7 52.9%
- 3 Kimi K2.6 52.3%
- 4
DeepSeek-V4-Pro 52.2%
- 5
Seed2.0 Pro 50%
- 6
Gemini 3.5 Flash 49.6%
- 7
Qwen 3.6 Max (Preview) 47.4%
- 8 GLM-5.1 42.7%
- 9 Grok 4.3 40.6%
- 10 gpt-oss-safeguard-120b 23.6%
- 11 gpt-oss-safeguard-20b 22.5%
Source: Official SafePyramid leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 54.2% | Apr 2026 | Official leaderboard | |
| 2 | Claude Opus 4.7 | 52.9% | Apr 2026 | Official leaderboard | |
| 3 | Kimi K2.6 | 52.3% | Apr 2026 | Official leaderboard | |
| 4 | | 52.2% | Apr 2026 | Official leaderboard | |
| 5 | | 50% | Feb 2026 | Official leaderboard | |
| 6 | | 49.6% | May 2026 | Official leaderboard | |
| 7 | Hunyuan-HY3-Preview | 49.1% | Official leaderboard | ||
| 8 | | 47.4% | Apr 2026 | Official leaderboard | |
| 9 | GLM-5.1 | 42.7% | Apr 2026 | Official leaderboard | |
| 10 | Grok 4.3 | 40.6% | Apr 2026 | Official leaderboard | |
| 11 | gpt-oss-safeguard-120b | 23.6% | Oct 2025 | Official leaderboard | |
| 12 | gpt-oss-safeguard-20b | 22.5% | Oct 2025 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 54.2% | Apr 2026 | Official leaderboard | |
| 2 | Claude Opus 4.7 | 52.9% | Apr 2026 | Official leaderboard | |
| 3 | Kimi K2.6 | 52.3% | Apr 2026 | Official leaderboard | |
| 4 | | 52.2% | Apr 2026 | Official leaderboard | |
| 5 | | 50% | Feb 2026 | Official leaderboard | |
| 6 | | 49.6% | May 2026 | Official leaderboard | |
| 7 | | 47.4% | Apr 2026 | Official leaderboard | |
| 8 | GLM-5.1 | 42.7% | Apr 2026 | Official leaderboard | |
| 9 | Grok 4.3 | 40.6% | Apr 2026 | Official leaderboard | |
| 10 | gpt-oss-safeguard-120b | 23.6% | Oct 2025 | Official leaderboard | |
| 11 | gpt-oss-safeguard-20b | 22.5% | Oct 2025 | Official leaderboard |
Rankings by part
We read 3 parts from the benchmark’s own leaderboard: Level 0: individual rules, Level 1: rule dependencies and Level 2: novel frameworks. Each ranks models on one part of the test. Claude Opus 4.7 leads 1 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Level 0: individual rules
11 models we track have a score here.
- 1 Claude Opus 4.7 ↑1 73.8%
- 2
DeepSeek-V4-Pro ↑2 70.8%
- 3 GPT-5.5 ↓2 70.7%
- 4 Grok 4.3 ↑5 70.6%
- 5
Qwen 3.6 Max (Preview) ↑2 68.9%
- 6 Kimi K2.6 ↓3 68.3%
- 7 GLM-5.1 ↑1 67%
- 8
Seed2.0 Pro ↓3 63.1%
- 9
Gemini 3.5 Flash ↓3 61.4%
- 10 gpt-oss-safeguard-120b 56.8%
Level 1: rule dependencies
11 models we track have a score here.
- 1 GPT-5.5 56.8%
- 2
DeepSeek-V4-Pro ↑2 54.5%
- 3
Seed2.0 Pro ↑2 54.1%
- 4 Kimi K2.6 ↓1 54.1%
- 5
Gemini 3.5 Flash ↑1 53.9%
- 6 Claude Opus 4.7 ↓4 51.1%
- 7
Qwen 3.6 Max (Preview) 46.5%
- 8 GLM-5.1 41.3%
- 9 Grok 4.3 39.4%
- 10 gpt-oss-safeguard-120b 12.1%
Level 2: novel frameworks
11 models we track have a score here.
- 1
Gemini 3.5 Flash ↑5 33.5%
- 2 Kimi K2.6 ↑1 33.1%
- 3
Seed2.0 Pro ↑2 32.9%
- 4 GPT-5.5 ↓3 32.9%
- 5 Claude Opus 4.7 ↓3 31%
- 6
DeepSeek-V4-Pro ↓2 30.6%
- 7
Qwen 3.6 Max (Preview) 26.4%
- 8 GLM-5.1 17%
- 9 Grok 4.3 11.9%
- 10 gpt-oss-safeguard-120b 2.10%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official SafePyramid leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official SafePyramid leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More 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%
- 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.
Gemini 4 Carbon 0%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 6 months ago, in May 2026, for Gemini 3.5 Flash. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the official leaderboard (12 entries).
- Pass Models scored 11 models we track, plus 1 other entry on its leaderboard.
- Warning Newest score The newest came 6 months ago, in May 2026, for Gemini 3.5 Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
- Pass Headroom The best is 54.2%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on bytedance.github.io: 11 rows for models we track and 1 other entry at the last read.
- Pass Last read November 18th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.