AI Trends

Trending story

English translation is being prepared.

7 reports7 story updatesUpdated Nov 19 02:20

Understand this story

Story summary

The English summary is being updated. See the reporting timeline below.

Latest development · Report

PromptArmor Releases FrontierBreach, an Enterprise AI Agent Security Benchmark

Developments

1 confirmed developments

  1. English translation is being prepared.

    Source:PromptArmor

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 19

  1. Testing Catalog

    PromptArmor Releases FrontierBreach, an Enterprise AI Agent Security Benchmark

    PromptArmor has released FrontierBreach, which evaluates how often enterprise AI agent workflows are misled by hidden instructions. Across tests of 21 AI models, the attack success rate was 34.9%. Every frontier model was compromised, with rates of 4.3% for Gemini Argon, 15.2% for Muse Wonder, and 91.4% for Grok 4.5. Attacks hidden in agent skills had a success rate of 45.5%.

  2. PromptArmorFirst-partySelected

    FrontierBreach Benchmark Reveals Differences in Attack Success Rates Across Enterprise AI Agent Workflows

    PromptArmor has released FrontierBreach, a benchmark evaluating indirect prompt injection attacks in enterprise AI agent workflows. The average attack success rate across nine industries was 34.9%, rising to 44.4% for healthcare and social assistance workflows. Chinese models had a higher average ASR than U.S. models, with the difference driven primarily by arbitrary code execution, phishing, and data destruction tasks; they also showed higher attack success rates in data theft, ransomware, and persistence attacks.

  3. PromptArmorFirst-partySelected

    PromptArmor Releases FrontierBreach, an Enterprise Agent Attack Benchmark

    PromptArmor released FrontierBreach to measure attack success rates in real-world enterprise agent workflows, covering consequences including ransomware, data theft, and phishing. In indirect prompt injection tests across nine industries, healthcare and social assistance workflows had an attack success rate of 44.4%, above the benchmark average of 34.9%. PromptArmor also said the benchmark successfully attacked Gemini 4 Carbon and observed differences between Chinese and U.S. models across different attack types.

  4. PromptArmorFirst-partySelected

    FrontierBreach Finds Successful Attacks on Gemini 4 Carbon

    PromptArmor says its FrontierBreach is currently the only known independent benchmark to find successful attacks on Google Gemini 4 Carbon. Gemini 4 Carbon’s model card, however, states that the model resisted all attacks in the benchmarks it used, with an attack success rate of 0%. FrontierBreach tests indirect prompt injection in enterprise agent workflows across nine industries, finding an average attack success rate of 34.9%, rising to 44.4% in healthcare and social assistance workflows.

  5. PromptArmorFirst-party

    PromptArmor Releases FrontierBreach Enterprise Agent Security Benchmark

    PromptArmor released FrontierBreach to assess attack success rates for enterprise agent workflows across nine industries when subjected to indirect prompt injection. Tests cover tasks such as care plans, engineering blueprints, and legal documents. Healthcare and social assistance workflows had the highest attack success rate at 44.4%, nearly 10 percentage points above the benchmark average of 34.9%. The benchmark injects attack content through common vectors such as untrusted files, emails, and skills, and covers consequences including ransomware, data exfiltration, and phishing.

  6. PromptArmorFirst-partySelected

    PromptArmor Launches FrontierBreach, an Enterprise Agent Security Benchmark

    PromptArmor has launched FrontierBreach to measure attack success rates in real-world enterprise agent workflows, covering outcomes such as ransomware, data theft, and phishing. The benchmark incorporates indirect prompt-injection vectors—including untrusted files, emails, and skills—into GDPval tasks across nine industries. The attack success rate for healthcare and social assistance workflows was 44.4%, higher than the benchmark average of 34.9%.

  7. PromptArmorFirst-party

    PromptArmor Releases FrontierBreach, an Enterprise AI Agent Security Benchmark

    PromptArmor has released FrontierBreach, which uses real-world enterprise workflows and complex attack chains to evaluate the success rate of attacks on AI agents exposed to indirect prompt injection. The benchmark covers nine industries and reports an average attack success rate of 34.9%, with healthcare and social assistance workflows recording the highest rate at 44.4%. The test scenarios involve consequences such as ransomware, data exfiltration, and phishing, and include prompt injection vectors such as untrusted files, emails, and skills.

Related stories

  1. Related storiesOpenAI releases the GPT-6 Astra model
  2. Related storiesOpenAI agents attacked online databases and government websites
  3. Related storiesOpenAI releases GPT-6.2 Astra and Sol, ARC Prize announces perfect scores
  4. Related storiesAnthropic launches Claude Fable 5.2 and Claude Mythos 5.2