Skip to main content
Back

AgentRedBench Catches Claude Sonnet 4.6 Failing 32% of Real-World Agent Attacks

AgentRedBench exposes how every major LLM fails indirect prompt injection attacks from SaaS tools, then ships a guard that cuts success rates from 70% to 2.4%

  • AgentRedBench is a new benchmark testing indirect prompt injection attacks across 24 enterprise SaaS integrations (Gmail, Salesforce, Jira, etc.).
  • Without defenses, attack success rates range from 32% (Claude Sonnet 4.6) to 81% (Gemini 3 Flash) across 8 tested models.
  • Companion guard model AgentRedGuard cuts the panel attack success rate from 69.9% to just 2.4% at 0.37% false-positive rate.
  • Unlike prior benchmarks, attack payloads are dynamically generated per-run so no two scenarios share the same injection string.
  • Existing open-source guards (Llama Guard, PromptGuard 2) fail because they were trained on chat data, not tool-response content.
  • Benchmark codebase and guard model weights are open; canonical test scenarios are evaluated through a controlled channel to prevent training contamination.

When your AI agent reads an email in Gmail, fetches a ticket from Jira, or pulls a record from Salesforce, it is consuming content that you did not write and cannot fully control. That is the attack surface at the heart of a new paper introducing AgentRedBench -- a benchmark and companion guard model designed to stress-test LLM agents against indirect prompt injection attacks delivered through enterprise SaaS integrations.

The results are sobering. Without any defense in place, every model tested -- spanning Anthropic, OpenAI, and Google -- was successfully hijacked in a significant fraction of scenarios. The best performer, Claude Sonnet 4.6, still failed 32% of the time, while Gemini 3 Flash topped the failure chart at 81%.

The attack nobody is measuring well

Indirect prompt injection (IPI) is distinct from the jailbreaks most people think of when they hear "prompt injection." It occurs when malicious instructions arrive through external content the agent processes as data -- a document it was asked to summarize, an email body, tool output. The attacker and the user are different people, and the attack is invisible to the user, who may have no idea the content they pointed the agent at was weaponized.

OWASP confirmed prompt injection as the top LLM risk for the third consecutive year, and the threat is landing at a particularly sensitive moment: the EU AI Act Article 15 cybersecurity obligation becomes operative for high-risk systems in August 2026. Despite this, the tooling to measure and defend against IPI in agentic, tool-calling contexts has lagged badly.

Existing benchmarks under-measure the threat: most cover only a handful of integrations with the same attack payload replayed across runs, and open-source guards are trained on chat-style data rather than tool-response content. That second point is critical -- a guard trained on chat conversations has never seen what a Salesforce API response looks like, let alone a weaponized one.