Malicious agentic influence campaigns (Anthropic)
Only in system cards
Agents receive requests to run multi-step influence operations with tools. Score is the refusal rate.
In-house Anthropic evaluation. The tasks are not public.
Made by Anthropic Released Apr 2026 Refusing harmful requests Tasks kept private Last in a system card Sep 2026
Last model added Sep 2026
- Level
- Everyday Most adults, without training
- Scored
- Graded by a model A language model scores each response, usually against a rubric.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
8
System cards report it, from 1 lab
No scores here yet
No lab, leaderboard or Epoch AI run has reported a score for Malicious agentic influence campaigns (Anthropic) yet.
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.
8 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.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 22 Sep 2026 | Claude Opus 5.5 system card | Agentic influence operation evaluation, helpful-only variant, median campaign execution score out of 32 at the highest difficulty: voter suppression 27.5, domestic polarization 28, no overall (Table 5.1.3.A) | Persuasion | |
| 1 Sep 2026 | Claude Fable 5.1 system card | Agentic influence operation evaluation, helpful-only variant, median campaign execution score out of 32 at the highest difficulty: voter suppression 29.5, domestic polarization 28.5, no overall (Table 5.1.3.A) | Persuasion | |
| 24 Jul 2026 | Claude Opus 5 system card | Agentic influence operation evaluation (updated), helpful-only variant, median campaign execution score out of 32 at the highest difficulty over 8 attempts: voter suppression 23.5, domestic polarization 26.0, no overall (Table 5.1.3.A) | Persuasion | |
| 30 Jun 2026 | Claude Sonnet 5 system card | Agentic influence operation evaluation, helpful-only variant, task completion rate out of 70 criteria: voter suppression 50.8%, domestic polarization 43.3%, no overall (Table 5.1.3.A) | Persuasion | |
| 9 Jun 2026 | Claude Fable 5 system card | Agentic influence operation evaluation, helpful-only Mythos 5 variant, average share of 70 success criteria completed: voter suppression 67.1%, domestic polarization 46.8%, no overall (Table 5.1.3.A) | Persuasion | |
| 28 May 2026 | Claude Opus 4.8 system card | Agentic influence operation evaluation, helpful-only variant, average share of 70 success criteria completed, no overall: voter suppression 73.3%, domestic polarization 55.1% (Table 5.1.3.A) | Persuasion | |
| 16 Apr 2026 | Claude Opus 4.7 system card | Agentic influence operation evaluation, helpful-only model, average task completion rate out of 70 criteria, no overall: voter suppression 57.1, domestic polarization 46.8. Final model near 0% (Table 5.1.3.A) | Persuasion | |
| 7 Apr 2026 | Claude Mythos Preview system card | Malicious agentic influence campaigns, helpful-only model task completion rate (share of 70 success criteria), voter suppression 59.5 and domestic polarization 42.1, no overall. Text says the released model completed near 0% because it refused (Table 8.3.1.3.A) | Safety |
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% - 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%