Published
- 5 min read
By Allan D - Editor, AI Security Wire
Check Point Argues Weak LLM Guardrails Lose to Stronger Models (PuzzleMask)
Check Point Research published an essay on 19 November 2026 with a blunt claim for anyone running a small LLM in front of a bigger one: the guard is structurally disadvantaged. The paper, “30 Years of Gatekeeper Asymmetry Doom: The Paradox that Makes Powerful LLMs More Vulnerable,” argues that filtering input with something less capable than the thing that consumes it is a losing structure, and that LLM pipelines are the newest example.
This is a vendor essay, not a peer-reviewed result, and the evidence for the AI section is Check Point’s own experiment. We report it as such. The essay also includes what it calls a pitch for why a security vendor should be “AI first”, and says LLMs can narrow these gaps, in some cases to nothing.
The principle
Check Point states it this way: if the system vetting adversarial input has strictly less power than the system processing that input, without leveraging some hard problem, this will inevitably lead to doom. The exception it carves out is signature verification, which is far simpler than running a file but can still, at least in a certain theoretical world, determine the answer.
The essay walks through older cases where the gap decided the outcome:
- Hash and pattern matching. A hash list is trivially defeated by changing a non-load-bearing byte. Regular expressions cannot capture assembly routines that have been moderately mutated.
- Deny lists. Check Point points to Log4Shell, where a deny list for
jndi:ldapwould have missed equally valid obfuscated exploit strings that were circulating in the wild. - Sandboxes. Any capability a real victim machine has and the sandbox lacks becomes an evasion. The example is mouse movement from a real human, which the authors say at the time evaded every state-of-the-art sandbox they tried in earlier research, and which they mitigated with simulated movement rather than publishing the evasion.
- Malicious Office documents. Defenders used OCR to flag images saying “please enable macros.” Attackers added adversarial noise that confused OCR libraries but not human readers, until Microsoft announced on 7 February 2022 that it would block internet macros by default (rollout began in April 2022).
The essay also gives a formal intuition: under the Time Hierarchy Theorem, a target that can run longer can compute classifications that a time-limited gatekeeper provably cannot. Check Point says the mapping from the theorem to malware classification, such as whether a binary edits the registry, is illustrative rather than a formal proof.
PuzzleMask
The AI case is Check Point’s own work. In late 2025 it tested a common setup: a quick LLM checks input before it reaches a heavier model that has high reasoning effort and a code interpreter.
PuzzleMask hides a policy-violating payload inside an obfuscated wrapper. The small checker fails to notice the payload and passes the prompt on. The larger model notices that the payload exists, extracts it and treats it as further input. Check Point calls this using the target’s intelligence against it. The technique is purely linguistic, with no emojis, invisible fonts or formatting tricks, and the wrapper can take arbitrary form, such as furniture assembly instructions or a credits reel.
Two limits matter. First, Check Point says PuzzleMask is not a jailbreak: it gets a prompt past an external screen without review, and says nothing about whether the target model will then comply. It can be combined with a jailbreak by using one as the payload. Second, it targets an external guardrail, which Check Point distinguishes from earlier work by Bisconti et al. using adversarial poems against a model’s internal guardrails (arXiv 2511.15304).
Check Point says the attack worked on GPT-5.2-thinking-high and was blocked by Opus 4.5, after Anthropic introduced an enhanced guardrail on its Claude Opus models. It says that guardrail does not entirely close the gap while the model remains more powerful than the guard, and links this to Anthropic’s later self-assessment layer. Those statements about vendor behavior come from Check Point alone, and we have not confirmed them with either vendor. The essay does not publish the full prompt, only a redacted illustration, and gives no trial counts or success rates and no OpenAI response.
What the authors recommend
Check Point does not conclude that gatekeepers are pointless. It calls doom a property of perfect play and notes real attackers are rarely perfect, citing the SolarWinds operation, which Check Point says shipped a consumer-grade Cobalt Strike payload. It adds that if a lot of malware could bypass a mitigation but does not, that is still a win.
Its questions for any attack surface:
- Is this surface doomed?
- Is there a feasible non-doomed gatekeeper to pivot to?
- How doomed is it in practice, and does the exploit exist in the wild now or only in a whitepaper?
- How much can the window be narrowed by increasing the gatekeeper’s power?
- What is the defense-in-depth plan when it is breached?
Defensive reading
Our assessment, not Check Point’s: the practical lesson for teams building LLM pipelines is about placement. A cheap classifier in front of a capable agent is a speed bump. It can reduce noise from unsophisticated input, but should not be the control that stops a destructive action.
- Enforce limits where the action happens, with permissions, allow-lists for tools and file paths, and human confirmation for irreversible steps, rather than relying on input screening alone.
- If screening is needed, give the guard capability close to the target’s, including the same reasoning budget and tool access, and accept the cost.
- Log what the guard passes as well as what it blocks, so a bypass leaves a trail.
- Treat Check Point’s model results as a single-vendor report until they are reproduced.
No evidence found as of 20 November 2026 in the Check Point post of PuzzleMask being used in the wild; we searched no other source. The essay describes it as a research experiment.
References
Frequently Asked Questions
- What is gatekeeper asymmetry doom?
- It is a principle Check Point Research states as: if the system vetting adversarial input has strictly less power than the system processing that input, and does not leverage some hard problem, this will inevitably lead to doom. In effect, a filter that can compute less than its target can be fooled by input the target understands and the filter does not.
- What is PuzzleMask?
- PuzzleMask is a prompt-crafting technique Check Point says it developed in late 2025. A payload is wrapped in obfuscated text so a small, fast LLM sanity check classifies the prompt as benign, while a larger model with high reasoning effort and a code interpreter can notice the payload, extract it and treat it as further input. Check Point says it is not a jailbreak and makes no claim that it makes a target model more likely to disobey its own guardrails.
- Does this mean guardrails are useless?
- Check Point says no. It describes the doom as a property of perfect play and argues that most real attackers are not playing perfectly, so narrowing the gap between gatekeeper and target still blocks real attacks. Its suggested questions include whether a non-doomed gatekeeper exists, how close the exploit is in the wild, and what the defense-in-depth plan is once the gate is breached.