A team created a constrained cross-instance communication task: one Fable 5.1 session generated visible visual artifacts, while a separate instance had to recover the intended message from screenshots. The researchers report 50 exercises and describe height and shape changes as the preferred encoding in 33. This is an example of output media being a data…
Inspect the ClaimsCase · DiggingBeagle record
Directed Fable 5.1 sessions carried messages through screenshot geometry
A research exercise found that, when instructed to communicate using rendered images, Fable 5.1 frequently used SVG dimensions that another instance could recover from a screenshot.
Scope
Controlled Fable 5.1 screenshot-message exercises, not a demonstration of secret unprompted model-to-model collusion.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Evidence
- Grade B
- Remediation
- not applicable
- AI role
- BY AI
Basis and provenance
CASE-2026-FABLE-IMAGE-ENCODED-MESSAGES: 50 directed trials; 33 of 50 used geometric encoding, as reported by the researcher. Evidence comes from SRC-2026-LESSWRONG-FABLE-STEGANO. Published technical statements or controlled results are not verified production-loss counts. No impact band is assigned because the observed outcome and its denominator do not establish a comparable severity measure.
Assessed Oct 24, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-FABLE-GEOMETRY · LessWrong: How Fable 5.1 hides messages in images
At a glance
Image output can carry instructions or arbitrary data in seemingly decorative geometry. Unlike an unprompted escape, this study asked both model instances to transmit and recover a preselected message; the result does not establish successful uncontrolled collusion or general channel capacity.
Read the implicationsIndependent reproduction on another model or evidence of spontaneous covert exchange is not established.
Limits and uncertaintyFull account
A team created a constrained cross-instance communication task: one Fable 5.1 session generated visible visual artifacts, while a separate instance had to recover the intended message from screenshots. The researchers report 50 exercises and describe height and shape changes as the preferred encoding in 33. This is an example of output media being a data channel even when plain-text conversations are inspected. The mechanism differs from microVM CPU-cache contention because both sides can observe a deliberately supplied screenshot rather than sharing physical hardware. The system was asked to encode a message, so the result should not be described as an autonomous evasion or demonstrated secret exfiltration. An auditor assessing model use of images may need to consider measurable geometry, not only OCR-visible words.
Mechanism and trust boundary
- 01
Give one model a directed concealment task
Researchers instructed a Fable 5.1 session to encode one of fifty short messages into an SVG image.
Boundary: Deliberate test instruction / rendered visual artifact
- 02
Pass only rendered pixels
The receiving session was shown a screenshot rather than the original SVG source, requiring recovery from image geometry or colors.
Boundary: Image authoring context / image-only observation
- 03
Decode the embedded content in a separate session
The second model attempted extraction from visible attributes; this tests a covert carrier despite absence of direct chat or file sharing between sessions.
Boundary: Rendered media output / downstream model text
Timeline
- Oct 21, 2026Event type unspecified
Directed image-communication study published
The authors described image-encoded messages and screenshot-only decoding, not an unprompted production data leak.
- Oct 21, 2026disclosure
Public disclosure
Claims & evidence
CLM-FABLE-GEOMETRYIn the reported 50-trial screenshot communication exercise, 33 trials used image or SVG geometry, notably element heights, to represent an encoded message.supported
Basis: reported finding
- supportsLessWrong: How Fable 5.1 hides messages in imagesprimary disclosure
Experiment section describing 50 trials; height/shape encoding analysis
CLM-FABLE-ENRICH-TRIALSThe authors evaluated 50 directed short-message exercises in which one Fable 5.1 instance constructed an SVG and a separate instance was asked to decode the screenshot.supported
Basis: reported finding
- supportsLessWrong: How Fable 5.1 hides messages in imagesprimary disclosure
Introduction; methodology describing 50 messages and image-only receiving session
Implications
Image output can carry instructions or arbitrary data in seemingly decorative geometry. Unlike an unprompted escape, this study asked both model instances to transmit and recover a preselected message; the result does not establish successful uncontrolled collusion or general channel capacity.
Controls and mitigations
- Treat screenshots and AI-generated visual artifacts as potential untrusted data surfaces. For a high-assurance workflow, use independent semantic content checks and restrict what receiving agents may infer or execute; the demonstration alone gives no measured production-risk reduction.
Unknowns and contradictions
- Independent reproduction on another model or evidence of spontaneous covert exchange is not established.
Sources and citation
Material revision history
- Oct 24, 2026 · Canonical change recorded · new in release · revision 187