Benchmark Almanac

MCP-SafetyBench

o4-mini 21.2% leads since December 2025
Not in model cards

Agents complete tool-use tasks across five domains. The benchmark reports both task completion and resistance to adversarial inputs.

The December 2025 paper compares GPT, Claude, Gemini and open models. OpenAI o4-mini achieved 21.22% overall task success and Qwen3-235B recorded 29.80% attack success.

Released
December 2025
Kind
Jailbreaks and prompt injection
Level
University
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Propensity
Language
English
In and out
Text → Actions, Text
Interaction
Agentic
Environment
APIs and tools
Refresh
One fixed set
Run by
Anyone
In views
Agents
Strongest model tested
GPT-5 #71 on the capability index
Last model added
September 2025
Scores refreshed
November 20th, 2026
13
Entries in its paper, 10 models we track

o4-mini leads with 21.2% among models released by September 2025 o4-mini leads with 21.2% among models released by September 2025

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. 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 September 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 o4-mini 21.2%
  2. 2 Gemini-2.5-Pro (snapshot unspecified) 20.4%
  3. 3 DeepSeek-V3.1 19.6%
  4. 4 GLM-4.5 18.4%
  5. 5 Grok 4 15.9%
  6. 6 GPT-5 15.9%
  7. 7 Claude 3.7 Sonnet 15.1%
  8. 8 Gemini 2.5 Flash Sep 2025 15.1%
  9. 9 Kimi K2 Sep 2025 14.3%
  10. 10 Qwen3-235B 10.2%
  11. 11 Claude Sonnet 4 10.2%
  12. 12 GPT-4.1 9.80%
  13. 13 GPT-4o (snapshot unspecified) 8.98%

Source: The MCP-SafetyBench paper

#EntryScoreSettingDateSource
1 o4-mini 21.2% Apr 2025 Official leaderboard
2 Gemini-2.5-Pro (snapshot unspecified) 20.4% Official leaderboard
3 DeepSeek-V3.1 19.6% Aug 2025 Official leaderboard
4 GLM-4.5 18.4% Aug 2025 Official leaderboard
5 Grok 4 15.9% Jul 2025 Official leaderboard
6 GPT-5 15.9% Aug 2025 Official leaderboard
7 Claude 3.7 Sonnet 15.1% Feb 2025 Official leaderboard
8 Gemini 2.5 Flash Sep 2025 15.1% Sep 2025 Official leaderboard
9 Kimi K2 Sep 2025 14.3% Sep 2025 Official leaderboard
10 Qwen3-235B 10.2% Official leaderboard
11 Claude Sonnet 4 10.2% May 2025 Official leaderboard
12 GPT-4.1 9.80% Apr 2025 Official leaderboard
13 GPT-4o (snapshot unspecified) 8.98% Official leaderboard
#ModelScoreSettingDateSource
1 o4-mini 21.2% Apr 2025 Official leaderboard
2 DeepSeek-V3.1 19.6% Aug 2025 Official leaderboard
3 GLM-4.5 18.4% Aug 2025 Official leaderboard
4 Grok 4 15.9% Jul 2025 Official leaderboard
5 GPT-5 15.9% Aug 2025 Official leaderboard
6 Claude 3.7 Sonnet 15.1% Feb 2025 Official leaderboard
7 Gemini 2.5 Flash Sep 2025 15.1% Sep 2025 Official leaderboard
8 Kimi K2 Sep 2025 14.3% Sep 2025 Official leaderboard
9 Claude Sonnet 4 10.2% May 2025 Official leaderboard
10 GPT-4.1 9.80% Apr 2025 Official leaderboard

Rankings by part

We read 10 parts from the benchmark’s paper in 2 groups: task success and attack success. Each ranks models on one part of the test. GPT-5 leads 3 of them, and 7 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Location Navigation

10 models we track have a score here.

  1. 1 o4-mini 18.9%
  2. 2 DeepSeek-V3.1 15.1%
  3. 3 Claude 3.7 Sonnet ↑3 13.2%
  4. 4 Grok 4 13.2%
  5. 5 GPT-4.1 ↑5 9.43%
  6. 6 GLM-4.5 ↓3 9.43%
  7. 7 Kimi K2 Sep 2025 ↑1 9.43%
  8. 8 Gemini 2.5 Flash Sep 2025 ↓1 9.43%
  9. 9 GPT-5 ↓4 5.66%
  10. 10 Claude Sonnet 4 ↓1 1.89%

Number one over time

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

OpenAI
Benchmark paper Not tested since Dec 2025
o4-mini 21.2%

Source: The MCP-SafetyBench paper

Scores over time

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

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) o4-mini Claude 3.7 Sonnet 0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) o4-mini

Source: Official MCP-SafetyBench 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

All 636 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 14 months ago, in September 2025, for Gemini 2.5 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 benchmark’s paper (13 entries).
  • Pass Models scored 10 models we track, plus 3 other entries on its leaderboard.
  • Warning Newest score The newest came 14 months ago, in September 2025, for Gemini 2.5 Flash.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5 (#71).
  • Pass Headroom The best is 21.2%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 10 rows for models we track and 3 other entries at the last read.
  • Pass Last read November 20th, 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.

How we check