←
Shankar Krishnan77 posts
Shankar Krishnan avatar

Shankar Krishnan

@PromptArmor

promptarmor.com

◷ Joined November 2023▧ 28 Photos and videos
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·1h

Today we’re releasing FrontierBreach, a new standard for attack success rate on agentic workflows in the enterprise. Existing benchmarks systematically underrepresent the security risks facing modern AI agents. FrontierBreach reflects attack scenarios on realistic enterprise workflows, while using sophisticated attack chains that result in cyber consequences (ransomware, data exfiltration, phishing, etc).

Post attachment
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·1h

Curiously, models from China have a higher ASR than models from the US on average, which was disproportionately due to their performance against arbitrary code execution, phishing, and data destruction tasks. These models performed better than their US counterparts on Data Theft, Ransomware, and Persistence attacks.

Post attachment
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·1h

We are also the ONLY known independent benchmark to have found successful attacks on Google's Gemini 4 Carbon. Even Carbon's own model card states "Gemini 4 Carbon resisted all attacks to achieve a 0% attack success rate" with the benchmark they used.

Post attachment
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·1h

Across nine industries, we tested indirect prompt injection during the kinds of work enterprises delegate to agents, from preparing nursing care plans to producing engineering blueprints and drafting legal briefs. Healthcare and social assistance workflows had the highest attack success rate in FrontierBreach at 44.4%, nearly 10 percentage points above the benchmark average of 34.9%.

Post attachment
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·1h

We wanted tasks where agents must navigate adversarial content while attempting to produce high-value deliverables. To do this, we have pioneered both the concept of an "Attack Soul" and a synthetic injection generation technique that can be utilized with any existing evaluation benchmark. For our flagship benchmark, we injected common delivery vectors for prompt injections (such as untrusted files, emails, and skills) for agent tasks from GDPval across all industry sectors.

0:13
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·Oct 28

We tested across multiple cyber risks, including phishing, data exfiltration and propensity to run malware. The differences are shocking!

Post attachment
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·Oct 28

Are models from China are more likely to exfiltrate data? PromptArmor tested all of our most viral attacks on a mix of models, and found that Deepseek, Kimi, GLM, and Qwen models are 88% more susceptible to cyber attacks than GPT Claude and Gemini models.

Post attachment
Snowflake avatar
⟳ Shankar Krishnan retweeted
Snowflake@Snowflake·Sep 8

Agent security gets harder the moment the model can act. Tool use, trust boundaries, delegated access, unintended side effects. That’s the real attack surface. Snowflake is working with PromptArmor to pressure-test new AI and agentic capabilities against those risks. Details: 👇 bit.ly/4xMW1L0

Post attachment
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·Jan 8

IBM AI prompt injected to execute malware! Also top of Hackernews right now, link in the comments:

Post attachment
Shankar Krishnan avatar
Shankar Krishnan@PromptArmor·26 Nov 2025

Excel files can be leaked by Claude AI! Quick action by Anthropic to mitigate this indirect prompt injection attack. Our coverage in The Information and full attack chain, below:

Post attachment