Skip to content
The Injection_
LIVE
LIVElast sweep next in--:--:--

[ RSS ][ JSON ][ llms.txt ]

PuzzleMask: a weak LLM guard misses what the strong model runs

PuzzleMask is a prompt-crafting technique that slips a hidden payload past a lightweight LLM guard and into the more capable model behind it.

· Check Point Research · Research by Check Point Research (Check Point Software Technologies)

GPT-5.2
MEDAttack

PuzzleMask hides a policy-violating payload in a wrapper that a lightweight LLM screen cannot read, while the stronger model behind it still acts on the payload. Check Point Research reports the technique worked on GPT-5.2-thinking-high.

01What happened

Check Point Research published PuzzleMask on November 19, 2026 inside a longer argument it calls gatekeeper asymmetry doom. The post adapts Moxie Marlinspike's Cryptographic Doom Principle to AI: when the system vetting adversarial input is strictly less powerful than the system processing it, and no hard problem is in play, the gatekeeper will eventually be bypassed.

02How it works

PuzzleMask wraps a policy-violating payload in an obfuscated form that a quick, lightweight LLM sanity check cannot resolve. The target model behind that screen runs with high reasoning effort and code-interpreter access, so it unpicks the wrapper, notices the hidden payload and acts on it. Check Point stresses that PuzzleMask is not itself a jailbreak and can be combined with one, and that the hidden payload can be arbitrary text.

03Who is affected

Deployments that put a small, fast LLM screen in front of a much more capable model are exposed, which is the common pattern for guardrails and content filters. Check Point reports the attack landed on GPT-5.2-thinking-high. Claude Opus 4.5 blocked it, which Check Point attributes to an enhanced guardrail Anthropic added to Opus models in late 2025.

04What to do

Check the power gap in each of your filtering layers: where the gatekeeper model is strictly weaker than the model it protects, Check Point advises raising the gatekeeper's capability or using the one you have more fully. The post also recommends pairing LLM screens with deterministic tools where the model is weak, and planning defense in depth for when a screen is bypassed.

Questions people ask

Is PuzzleMask a jailbreak?
No. Check Point Research states that PuzzleMask is not itself a jailbreak, and describes it instead as a prompt-crafting technique for bypassing quick LLM-based sanity checks. PuzzleMask hides what a payload is from the screening model, rather than persuading the target model to abandon its own policy. Check Point notes that the two can be combined.
Which models did Check Point test PuzzleMask against?
Check Point Research names two models in the post. The attack landed on GPT-5.2-thinking-high, a configuration with high reasoning effort and code-interpreter access. Claude Opus 4.5 blocked it, which Check Point attributes to an enhanced guardrail that Anthropic introduced in Opus models in late 2025. The post publishes no detection rates or benchmark tables for the technique.
How can defenders tell whether their own guardrail is exposed?
Check Point Research advises asking one question of each filtering layer: is the gatekeeper strictly less capable than the system it protects? Where the answer is yes, the post argues that a bypass will eventually exist. Check Point recommends raising the gatekeeper's capability, adding deterministic checks where the model is weak, and assuming the screen will be breached.

Sources 2 links, all fetched

  1. 1research.checkpoint.comCheck Point Research
  2. 2moxie.orgDoom Principle