Benchmark Almanac

FinRED Financial Safety

GPT-5 0.15% leads since June 2026
Not in model cards

Models answer prompts about financial risks. A finance-specific judge grades each response and reports the percentage that assist harmful conduct.

Selectstar and Korea’s Financial Security Institute built the test with 12 financial experts. The institute reports results for 12 target models across five risk areas.

Released
June 2026
Kind
Deception and alignment
Level
University
Scored
Graded by a model
Answers
Open-ended
Tasks
All public
Measures
Propensity
Behavior
Harmful use
In and out
Text → Text
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
Strongest model tested
GPT-5 #71 on the capability index
Last model added
August 2025
Scores refreshed
November 20th, 2026
12
Entries in its paper, 5 models we track

GPT-5 leads with 0.15% among models released by August 2025 GPT-5 leads with 0.15% among models released by August 2025

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The paper reports no overall attack success rate. The headline averages its five financial risk categories. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. 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 August 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 GPT-5 GPTFuzzer (Black-box) 0.15%
  2. 2 Gemini 2.5 Flash (snapshot unspecified) TAP (Black-box) 1.94%
  3. 3 GPT-5 mini GPTFuzzer (Black-box) 1.96%
  4. 4 Claude Sonnet 4 TAP (Black-box) 2.74%
  5. 5 Llama-3.1-8B-it TAP (Black-box) 3.81%
  6. 6 Gemini 2.5 Pro (snapshot unspecified) TAP (Black-box) 4.19%
  7. 7 Gemma 3 12B TAP (Black-box) 17.4%
  8. 8 qqWen-32B TAP (Black-box) 25.3%
  9. 9 EXAONE-3.5-7.8B-Instruct TAP (Black-box) 32.5%
  10. 10 Qwen2.5-7B-it TAP (Black-box) 33%
  11. 11 qqWen-7B Direct Request 34.7%
  12. 12 FinMA TAP (Black-box) 41.2%

Source: The FinRED Financial Safety paper

#EntryScoreSettingDateSource
1 GPT-5 0.15% GPTFuzzer (Black-box) Aug 2025 Official leaderboard
2 Gemini 2.5 Flash (snapshot unspecified) 1.94% TAP (Black-box) Official leaderboard
3 GPT-5 mini 1.96% GPTFuzzer (Black-box) Aug 2025 Official leaderboard
4 Claude Sonnet 4 2.74% TAP (Black-box) May 2025 Official leaderboard
5 Llama-3.1-8B-it 3.81% TAP (Black-box) Official leaderboard
6 Gemini 2.5 Pro (snapshot unspecified) 4.19% TAP (Black-box) Official leaderboard
7 Gemma 3 12B 17.4% TAP (Black-box) Mar 2025 Official leaderboard
8 qqWen-32B 25.3% TAP (Black-box) Official leaderboard
9 EXAONE-3.5-7.8B-Instruct 32.5% TAP (Black-box) Dec 2024 Official leaderboard
10 Qwen2.5-7B-it 33% TAP (Black-box) Official leaderboard
11 qqWen-7B 34.7% Direct Request Official leaderboard
12 FinMA 41.2% TAP (Black-box) Official leaderboard
#ModelScoreSettingDateSource
1 GPT-5 0.15% GPTFuzzer (Black-box) Aug 2025 Official leaderboard
2 GPT-5 mini 1.96% GPTFuzzer (Black-box) Aug 2025 Official leaderboard
3 Claude Sonnet 4 2.74% TAP (Black-box) May 2025 Official leaderboard
4 Gemma 3 12B 17.4% TAP (Black-box) Mar 2025 Official leaderboard
5 EXAONE-3.5-7.8B-Instruct 32.5% TAP (Black-box) Dec 2024 Official leaderboard

Rankings by part

We read 5 parts from the benchmark’s paper, each ranking models on one part of the test, such as R1: Cyber Threats, R2: Financial Crime and R3: Misinformation. GPT-5 leads 4 of them, and Claude Sonnet 4 leads the other one. Arrows show how far a model moved from its overall rank among the models we track.

R1: Cyber Threats

5 models we track have a score here.

  1. 1 GPT-5 0%
  2. 2 GPT-5 mini 0%
  3. 3 Claude Sonnet 4 4.43%
  4. 4 Gemma 3 12B 33.8%
  5. 5 EXAONE-3.5-7.8B-Instruct 75.3%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
EXAONE-3.5-7.8B-Instruct
TAP (Black-… AutoDAN (Wh… GPTFuzzer (…
89%
Gemma 3 12B
TAP (Black-… AutoDAN (Wh… GCG (White-… GPTFuzzer (…
67.6%
Claude Sonnet 4
TAP (Black-…
7.88%
GPT-5 mini
GPTFuzzer (…
7.65%
GPT-5
GPTFuzzer (…
4.62%

Source: The FinRED Financial Safety paper

ModelEffortHarnessScoreCost per taskOutput tokens
EXAONE-3.5-7.8B-Instruct GPTFuzzer (Black-box) 89%
Gemma 3 12B GPTFuzzer (Black-box) 67.6%
Gemma 3 12B AutoDAN-Turbo 43.2%
EXAONE-3.5-7.8B-Instruct AutoDAN (White-box) 42.2%
Gemma 3 12B GCG (White-box) 39.7%
EXAONE-3.5-7.8B-Instruct GCG (White-box) 39.4%
EXAONE-3.5-7.8B-Instruct Direct Request 38.5%
EXAONE-3.5-7.8B-Instruct AutoDAN-Turbo 38.2%
EXAONE-3.5-7.8B-Instruct TAP (Black-box) 32.5%
Gemma 3 12B AutoDAN (White-box) 30%
Gemma 3 12B Direct Request 21.7%
Gemma 3 12B TAP (Black-box) 17.4%
Claude Sonnet 4 Direct Request 7.88%
GPT-5 mini TAP (Black-box) 7.65%
Claude Sonnet 4 GPTFuzzer (Black-box) 5.30%
GPT-5 AutoDAN-Turbo 4.62%
GPT-5 mini AutoDAN-Turbo 4.26%
Claude Sonnet 4 AutoDAN-Turbo 4.19%
GPT-5 TAP (Black-box) 4.09%
GPT-5 mini Direct Request 3.47%
GPT-5 Direct Request 3.22%
Claude Sonnet 4 TAP (Black-box) 2.74%
GPT-5 mini GPTFuzzer (Black-box) 1.96%
GPT-5 GPTFuzzer (Black-box) 0.15%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
EXAONE-3.5-7.8B-Instruct
TAP (Black-… AutoDAN (Wh… GPTFuzzer (…
89%
Gemma 3 12B
TAP (Black-… AutoDAN (Wh… GCG (White-… GPTFuzzer (…
67.6%
Claude Sonnet 4
TAP (Black-…
7.88%
GPT-5 mini
GPTFuzzer (…
7.65%
GPT-5
GPTFuzzer (…
4.62%

Source: The FinRED Financial Safety paper

ModelEffortHarnessScoreCost per taskOutput tokens
EXAONE-3.5-7.8B-Instruct GPTFuzzer (Black-box) 89%
Gemma 3 12B GPTFuzzer (Black-box) 67.6%
Gemma 3 12B AutoDAN-Turbo 43.2%
EXAONE-3.5-7.8B-Instruct AutoDAN (White-box) 42.2%
Gemma 3 12B GCG (White-box) 39.7%
EXAONE-3.5-7.8B-Instruct GCG (White-box) 39.4%
EXAONE-3.5-7.8B-Instruct Direct Request 38.5%
EXAONE-3.5-7.8B-Instruct AutoDAN-Turbo 38.2%
EXAONE-3.5-7.8B-Instruct TAP (Black-box) 32.5%
Gemma 3 12B AutoDAN (White-box) 30%
Gemma 3 12B Direct Request 21.7%
Gemma 3 12B TAP (Black-box) 17.4%
Claude Sonnet 4 Direct Request 7.88%
GPT-5 mini TAP (Black-box) 7.65%
Claude Sonnet 4 GPTFuzzer (Black-box) 5.30%
GPT-5 AutoDAN-Turbo 4.62%
GPT-5 mini AutoDAN-Turbo 4.26%
Claude Sonnet 4 AutoDAN-Turbo 4.19%
GPT-5 TAP (Black-box) 4.09%
GPT-5 mini Direct Request 3.47%
GPT-5 Direct Request 3.22%
Claude Sonnet 4 TAP (Black-box) 2.74%
GPT-5 mini GPTFuzzer (Black-box) 1.96%
GPT-5 GPTFuzzer (Black-box) 0.15%

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 Jun 2026
GPT-5 0.15%

Source: The FinRED Financial Safety 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 GPT-5 Claude Sonnet 4 Claude Sonnet 4 Gemma 3 12B Gemma 3 12B Gemma 3 12B EXAONE-3.5-7.8B-Instruct 0% 20% 40% 60% 80% 100% 2025 GPT-5 Claude Sonnet 4

Source: Official FinRED Financial Safety 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

8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 15 months ago, in August 2025, for GPT-5. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
  • Pass Independent runs Scores from the benchmark’s paper (12 entries).
  • Pass Models scored 5 models we track, plus 7 other entries on its leaderboard.
  • Warning Newest score The newest came 15 months ago, in August 2025, for GPT-5.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5 (#71).
  • Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 24 rows for models we track and 7 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