Case · DiggingBeagle record

A poisoned model could embed inferred private attributes in ordinary text that a user later shares

Synthesis Smuggling is not a direct prompt instruction to leak stored secrets. The researchers modified model behavior with a LoRA-style poison so it estimated personal traits from a conversation, then embedded those guesses into generated content such as an email the user might copy to another recipient.

Public disclosure: Nov 12, 2026

Scope

An attacker can modify or distribute a model adapter used to generate text later shared by a user.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    An attacker can modify or distribute a model adapter used to generate text later shared by a user.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    Two encoding strategies used zero-width characters and a keyed lexical code. The covert channel can transfer the model’s inference without tool network access or an explicit disclosure in the visible answer. Attribution of the guessed traits may still be incorrect.

    Boundary: mechanism

  3. 03

    Observed consequence

    Model-inferred personal attributes become a covert payload in a user-authorized output artifact, beyond the visible answer and without a tool call.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-SYNTHESIS-MECHSynthesis Smuggling is not a direct prompt instruction to leak stored secrets. The researchers modified model behavior with a LoRA-style poison so it estimated personal traits from a conversation, then embedded those guesses into generated content such as an email the user might copy to another recipient. Two encoding strategies used zero-width characters and a keyed lexical code. The covert channel can transfer the model’s inference without tool network access or an explicit disclosure in the visible answer. Attribution of the guessed traits may still be incorrect.supported

Basis: reported finding

Link to claim
CLM-NOV15-SYNTHESIS-SCOPEThe study reports high decoder accuracy on controlled synthetic conversations with Qwen and Llama derivatives. It does not document a deployed checkpoint, real human attribute correctness, copied message reaching an attacker, or guaranteed survival after client formatting transformations. This is a model-artifact and output-provenance risk; it is different from TrojanStego demonstrations of memorized or observed secret transmission.supported

Basis: reported finding

Link to claim

Implications

Model-inferred personal attributes become a covert payload in a user-authorized output artifact, beyond the visible answer and without a tool call.

Controls and mitigations

  • Treat model adapter provenance and copied output as separate security boundaries; compare output text under normalized Unicode and format transformations.
  • Evaluate whether a model can infer attributes accurately before making any claim that real private facts were exposed; test with controlled, consented data.

Unknowns and contradictions

  • This is a model-artifact and output-provenance risk; it is different from TrojanStego demonstrations of memorized or observed secret transmission.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250