Case · DiggingBeagle record

Scaling an image could reveal attacker text to a vision-language model that was inconspicuous at the original resolution

The authors constructed images that looked largely ordinary at 512-pixel scale but made injected instructions more legible after downsampling to a 64-pixel input. The exploit depends on how the agent preprocesses pixels, not merely on whether a human would notice suspicious text in the source image.

Public disclosure: Nov 12, 2026

Scope

An agent retrieves or opens an attacker-influenced image and preprocesses it at a resolution different from the human-inspected artifact.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    An agent retrieves or opens an attacker-influenced image and preprocesses it at a resolution different from the human-inspected artifact.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    Across selected model tests, text recognition was much higher than the rate of successfully induced tool actions. In the web-agent evaluation, attacker-relevant behavior rose to 4.00 percent for Gemini-3-Flash and 3.67 percent for GPT-4o, despite much higher perception of injected instructions.

    Boundary: mechanism

  3. 03

    Observed consequence

    Untrusted image pixels cross into the model instruction channel when text only becomes apparent after resizing or enhancement.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-SISS-MECHThe authors constructed images that looked largely ordinary at 512-pixel scale but made injected instructions more legible after downsampling to a 64-pixel input. The exploit depends on how the agent preprocesses pixels, not merely on whether a human would notice suspicious text in the source image. Across selected model tests, text recognition was much higher than the rate of successfully induced tool actions. In the web-agent evaluation, attacker-relevant behavior rose to 4.00 percent for Gemini-3-Flash and 3.67 percent for GPT-4o, despite much higher perception of injected instructions.supported

Basis: reported finding

Link to claim
CLM-NOV15-SISS-SCOPEThe reported images are a controlled cross-resolution prompt-injection demonstration, not proof that most multimodal agents obey such instructions or that a deployed customer was compromised. The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.supported

Basis: reported finding

Link to claim

Implications

Untrusted image pixels cross into the model instruction channel when text only becomes apparent after resizing or enhancement.

Controls and mitigations

  • Inspect inputs at the actual model resolution and at multiple nearby scales before granting authority to extracted text.
  • Require external authorization for mutating tool calls regardless of whether an image instruction is recognized; evaluate real side effects separately from text-recognition rate.

Unknowns and contradictions

  • The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.

Sources and citation

Material revision history

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