Prime Intellect says an unauthorized party replaced a deployment behind its hosted DeepSeek V4.1 Flash endpoint. Requests reaching that deployment were answered by modified weights between November 8 at 01:24 UTC and November 10 at 06:35 UTC. The company published PI-SA-2026-001 on November 19.

The replacement was an abliterated model, a copy modified to reduce or remove refusal behavior. Prime Intellect says its ordinary capability tests stayed unchanged while refusals fell sharply. Customers whose workflows depended on those refusals should review affected outputs, the company advises.

The reported intrusion changed the model, not its public name

The notice identifies Prime Inference’s deepseek/deepseek-v4.1-flash endpoint. It says the intruder used publicly available modified weights while leaving the chat template, tokenizer and serving configuration unchanged. A stable endpoint name therefore did not ensure that requests reached the expected model artifact.

Evidence status: provider-reported unauthorized model substitution. This is an incident disclosure, rather than a controlled experiment by an outside researcher. The public notice gives no entry mechanism, attacker identity, affected request count or customer list. It says all impacted customers have been contacted.

The stated scope is one deployment during the specified window. The notice does not say that every request to the endpoint, every Prime Intellect model or the original model developer’s own service was affected.

Capability scores missed a different behavior change

Prime Intellect reports unchanged results on GPQA Diamond, a difficult question-answering test, and τ-bench, a test of agents handling tasks with tools. It separately reports these refusal counts:

Safety test Original model refusals Modified model refusals
JailbreakBench 98 of 100 prompts 1
StrongREJECT 304 of 313 prompts 2
HarmBench 198 of 240 prompts 0

These are the provider’s test results. They show why unchanged task performance can coexist with changed refusal behavior. They do not measure affected customer requests, successful attacks or harm caused during the incident.

For an agent, a model performing tasks with tools, refusal behavior is only one control. File access, command permissions and outbound connections should remain constrained by the surrounding software even if the model answers an unsafe request.

Separate model integrity from prompt confidentiality

Prime Intellect says it found no evidence of prompts or completions being recorded or leaving the serving cluster. It also reports no affected keys or customer credentials, no access beyond that cluster, and no modifications to other models.

The company says its endpoints use zero data retention, meaning requests are not logged. To support its assessment, it describes examining every recovered version of the attacker’s proxy, node disks and memory, and process-level session records. It says none reads or transfers request content.

Those statements describe the provider’s current investigation. The notice says the investigation continues and promises updates if its assessment changes. It does not publish the forensic records needed for an independent reconstruction.

Our canary-credential coverage concerns prompt exposure through processing routes. The two reports describe different failures. Neither source establishes that this substitution incident is one of the provider compromises in that research.

Review outputs and keep authorization outside the model

The supported immediate response is targeted:

  • Check whether the provider contacted your organization about the affected deployment. Use the exact endpoint and incident window when reviewing request records you already retain.
  • Review outputs and downstream actions where your workflow depended on refusals. Include any tool use those outputs authorized; generated text alone does not establish that an action occurred.
  • Keep access rules independent of model behavior. A changed model should still be unable to read unauthorized files or perform unapproved operations.
  • Ask the provider how model artifacts are approved and checked during deployment. The notice reports a replacement but leaves the entry path and preventive changes unspecified. It provides no customer-installed fixed version.

The disclosure does not support a blanket claim that customer credentials were stolen. Its confidentiality assessment and its confirmed integrity problem need separate follow-up.

What the thread adds

In the Hacker News discussion, platinumrad asks: “Is this an ad? Why are they breathlessly presenting benchmark results?” That skepticism concerns the provider’s framing, not independent evidence about the intrusion.

The benchmark table above is useful for a narrower reason: unchanged capability results did not validate the served model’s identity or refusal behavior. The discussion supplies no verified entry mechanism or additional victim evidence.

From a 4-comment thread, in HN’s own order.

Caveat: Prime Intellect operates the affected service and supplies both the forensic assessment and benchmark results; the public notice leaves the intrusion mechanism unresolved.