AI Guardrails Left to Developers, ClawSecure Report Finds

OpenAI said implementing guardrails is “the developer’s responsibility.” ClawSecure found 17 of 20 obfuscation techniques survive the MCP round trip intact.

SAN FRANCISCO, CA, UNITED STATES, October 13, 2026 /EINPresswire.com/ — OpenAI said implementing guardrails is “the developer’s responsibility” when ClawSecure reported prompt injection through OpenAI’s API, as documented in The AI Agent Threat Report, and Google told “all integrators building agents” to build their own sanitization. The two-volume report, which ClawSecure released on September 24, 2026, sets those statements beside the company’s own measurements of what AI guardrails are up against in practice, from disguised attack text to the models’ own judgment. For builders weighing guardrail products, the open question is where a guardrail has to sit to hold, and whether pattern matching or the model’s own judgment can do the job.

The measurements start with how attackers disguise their instructions. ClawSecure found that 17 of 20 obfuscation techniques, including zero-width Unicode, right-to-left overrides, homoglyphs, and fake system tags, survive the MCP round trip with attack markers intact, so pattern-matching defense does not hold this line. That finding concerns pattern-matching defense specifically, not every possible guardrail. The models’ own judgment did not close the gap either. ClawSecure tested 14 models from five frontier labs against one identical hardened attack, and no model defended cleanly. ClawSecure’s results ran from the best-defended model all the way to models that obeyed every time.

ClawSecure’s research concludes that security has to sit one layer below the AI model, inspecting the content an agent reads and every tool call it makes. AI guardrails in that position check what reaches the agent and what the agent tries to do, instead of relying on the model to refuse. Because no model defended cleanly, that conclusion applies across all five labs ClawSecure tested. For teams comparing AI guardrails, ClawSecure has published a guide to prompt injection defenses that hold up.

“The labs handed a security problem to 123 million people, and the only way to find that out is to file a disclosure and keep pressing until they answer,” said J.D. Salbego, Founder and CEO of ClawSecure.

For developers asking whether AI guardrails can stop prompt injection on their own, ClawSecure’s research concludes the fix cannot be the AI’s own judgment, because even the best-defended model it tested still obeyed hidden attacker commands 26.7% of the time. That result is why the research places security below the model rather than inside it. Readers can download both volumes at no cost from ClawSecure’s official release page. Every company named received coordinated disclosure before publication.

About ClawSecure
ClawSecure is the only end-to-end AI agent security platform that requires zero expertise, from free pre-deployment scans to runtime monitoring. Its flagship, the AI CISO™, is the first AI Chief Information Security Officer: it secures your system, handles every install and configuration, and manages your entire environment safely, your own security team without hiring one. ClawSecure is a member of the NVIDIA Inception Program, twice named #2 Product of the Day on Product Hunt, and selected from 30,000 AI startups for The Pitch by Deel (a16z, General Catalyst, J.P. Morgan).

Website: https://www.clawsecure.ai

Kevin Clinton
ClawSecure
+1 323-327-7528
Visit us on social media:
LinkedIn
YouTube
X

Media gallery