Cybersecurity Research

Why Weaker Gatekeepers Fail: Check Point Research Explains the 'Doom' Principle Behind Antivirus and AI Guardrails

Check Point Research · 19 Nov 2026
Key Takeaway Do not rely on a single simple filter, whether a signature scanner or an AI guardrail, to stop attacks that are more complex than the filter itself.

Check Point Research has set out a principle it calls "gatekeeper asymmetry doom". It says that if the system vetting hostile input has strictly less power than the system processing that input, and does not lean on some hard problem, failure is inevitable. The authors compare it to Moxie Marlinspike's Cryptographic Doom Principle, and joke that every quick sanity check involves either cryptography or a silent prayer.

The researchers say the idea explains several familiar security problems. They list the limits of static binary classification, the spread of sandbox evasions, and their own "PuzzleMask" research on bypassing external AI guardrails. They also point to the Time Hierarchy Theorem as a theoretical example of the same pattern. The article opens with the warm-up example that security practitioners know well.

In the 1990s, it was tempting to think malware was solved. Antivirus tools would compare each executed file against a database of known bad hashes. But the gatekeeper only checked whether a file was on a list, while the file itself could do almost anything a program can do. Attackers countered cheaply by changing a single unimportant byte in each build. Signature-based rules such as YARA, which match patterns like characteristic instruction sequences, are more powerful, and the authors assume a perfectly skilled rule writer to make the point. They still see a straight line from simple hashing to the long struggle of classifying files by static analysis alone.

Their recommendation is to spot "doomed" attack surfaces and narrow the gap, either by upgrading the gatekeeper or by using its existing power more fully. They say frontier LLMs, with scaffolding, extended reasoning and code execution, can help narrow these gaps, but remain subject to the same principle.

Why it matters

Security leaders who treat a pre-check, such as file signature scanning or an external AI guardrail, as the main barrier are exposed, because the check is weaker than what it guards. This breaks the assumption that a gatekeeper can fully judge what it screens, and it pushes teams to ask where a weak filter sits in front of a far more capable system. It also means adding an LLM as a gatekeeper does not escape the problem, since the source says LLMs are still subject to it.

What to do

  • Identify places where a simple filter screens input for a much more powerful system, such as file scanners or AI guardrails.
  • Treat hash-based and pattern-based detections as one layer, not as proof a file is safe.
  • Narrow the gap by upgrading gatekeeper capability or using its existing power more fully.
AI security LLM guardrails antivirus malware detection Check Point Research
Building or buying AI systems? Governing them under ISO 42001 ->

Summarised by CISO AI from Check Point Research, written with Claude Sonnet 5.5. We link back to every original so you can read it yourself.