Check Point sets out the gatekeeper asymmetry doom principle: when guardrails are weaker than the processing system, the attack surface is hard to defend
30 Years of Gatekeeper Asymmetry Doom: The Paradox that Makes Powerful LLMs More Vulnerable
Check Point Research proposes the "gatekeeper asymmetry doom" principle in information security: if a system that inspects adversarial input is strictly weaker in compute than the system that processes that input, defenses will eventually be bypassed. The article illustrates the principle with thirty years of cases, including hash comparison, YARA signature matching, static reverse engineering, denylists, sandbox escapes, and OCR attack and defense on "please enable macros" documents, and gives a mathematical argument via a time-hierarchy theorem. The author also discloses the PuzzleMask technique: wrapping a violating payload in an obfuscated plain-text shell so that a resource-constrained fast LLM security check fails to identify a problem and lets it through, after which a stronger target model parses out the payload; the technique itself is not a jailbreak but can be combined with jailbreak prompts; the author says that in experiments at the end of 2025 the attack worked on GPT-5.2-thinking-high but was blocked by the guardrails Anthropic introduced for Opus 4.5. The article recommends identifying attack surfaces destined to fall and narrowing the window of failure by upgrading the gatekeeper or making fuller use of its existing capabilities.