Prime Intellect reports unauthorized model swap on deepseek endpoint
An unauthorized party replaced the deepseek-v4.1-flash model with an abliterated version on Prime Intellect for two days in November 2026, removing safety refusals without affecting core capabilities.
Between November 8 and November 10, 2026, an unauthorized actor swapped the model behind Prime Intellect’s deepseek/deepseek-v4.1-flash endpoint with a modified version that bypasses safety refusals. The incident lasted approximately 53 hours, during which all requests to this specific endpoint were served by the altered model weights. Prime Intellect confirmed that no customer data was exfiltrated, but users relying on the model’s refusal behavior should review outputs from this window.
What happened
The incident began at 01:24 UTC on November 8, 2026, when an unauthorised party replaced the deployment behind the deepseek/deepseek-v4.1-flash endpoint. Instead of the standard model, the system served an "abliterated" copy. Abliteration refers to modifications made to a model’s weights to reduce or remove its tendency to refuse harmful or restricted prompts. The attacker used publicly available abliterated weights for this specific model version. The swap remained active until 06:35 UTC on November 10, 2026.
During the replacement, the chat template, tokenizer, and serving configuration remained unchanged. This means the interface and input processing looked identical to users, making the swap difficult to detect through normal interaction logs. Prime Intellect tested the modified weights and found that general capability benchmarks, such as GPQA Diamond and τ-bench, were unaffected. However, safety benchmarks showed a drastic reduction in refusals. On JailbreakBench, refusals dropped from 98 out of 100 prompts to just one. On StrongREJECT, they fell from 304 out of 313 to two, and on HarmBench, they went from 198 out of 240 to none.
Prime Intellect has contacted all customers impacted by the incident. They recommend that any application depending on the model’s refusal mechanisms review outputs generated between the start and end times of the breach. The company stated that no other models on their platform were modified, and no other clusters or infrastructure were accessed. Customer credentials and API keys remain secure.
How it works
The attack involved replacing the model weights file rather than compromising the underlying infrastructure or stealing data. By swapping the weights with a publicly available abliterated version, the attacker changed how the model responded to specific types of prompts without altering the serving code. This is a supply-chain style attack on the model artifact itself, rather than a traditional network intrusion. The term "abliterated" describes a process where specific safety-aligned behaviors are surgically removed from the neural network weights, often using techniques that target the vectors responsible for refusal responses.
Prime Intellect’s infrastructure uses Zero Data Retention (ZDR) for its endpoints, meaning it does not log requests or completions. This architectural choice limited the attacker’s ability to harvest data even if they had attempted to do so. The company recovered every version of the attacker's proxy and examined the disks and memory of every node involved. Their analysis of process-level records showed that the attacker’s sessions did not read or transfer request content. The investigation remains ongoing, but current evidence suggests the breach was confined to the model substitution.
Key details
- Timeline: The incident occurred from November 8, 2026, at 01:24 UTC to November 10, 2026, at 06:35 UTC.
- Affected Endpoint: Only the deepseek/deepseek-v4.1-flash endpoint was compromised; no other models or clusters were touched.
- Nature of Change: The model was replaced with publicly available abliterated weights, removing refusal behaviors while keeping capabilities intact.
- Data Security: No evidence was found that prompts or completions were recorded, stored, or left the serving cluster.
- Credentials: No API keys or customer credentials were affected or accessed.
- Detection: The chat template and tokenizer were unchanged, making the swap invisible to standard input/output checks.
Why it matters
This incident highlights a growing risk in AI infrastructure: the integrity of model weights. Developers often assume that if their API connection is secure, the model behavior is guaranteed. However, this event shows that an attacker can alter the model’s core behavior without breaking the API contract or stealing data. For applications that rely on safety refusals to prevent misuse, such as customer support bots or content moderation tools, this type of swap can lead to significant compliance and reputational risks. The fact that the capabilities remained unchanged makes it even harder to detect through performance monitoring alone.
For engineering teams, this underscores the importance of verifying model provenance and behavior, not just connectivity. Relying solely on the provider’s security is insufficient for high-stakes applications. Teams need to consider implementing output monitoring or hash verification for model weights if their use case is sensitive to safety alignment. The incident also demonstrates the value of Zero Data Retention policies. Because Prime Intellect did not log requests, the potential damage was limited to behavioral changes rather than data leakage, reducing the severity of the privacy impact.
What you can do
- Review all outputs from the deepseek/deepseek-v4.1-flash endpoint generated between November 8, 01:24 UTC and November 10, 06:35 UTC, 2026.
- Check if your application logic depends on the model refusing certain prompts and assess any potential harm from unrefused outputs during the incident window.
- Implement automated monitoring for unexpected changes in model refusal rates or tone, which can signal weight tampering.
- Verify the integrity of model weights if you host models locally, using cryptographic hashes to detect unauthorized modifications.
- Ensure your vendor contracts include clear protocols for notifying you of model swaps or behavioral changes.
- Consider adding a secondary safety layer in your application code that does not rely solely on the base model’s refusals.



