cd /news/ai-safety/prime-intellect-says-deepseek-v4-1-f… · home › topics › ai-safety › article
[ARTICLE · art-172905] src=primeintellect.ai ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

Prime Intellect says DeepSeek v4.1 Flash swap did not expose prompts

Prime Intellect disclosed that between November 8, 2026, at 01:24 UTC and November 10, 2026, at 06:35 UTC an unauthorised party replaced the deployment behind its deepseek/deepseek-v4.1-flash endpoint with publicly available abliterated weights, and the company said it found no evidence that prompts or completions were recorded or left its serving cluster. Prime Intellect reported that refusals on the modified weights fell from 98 of 100 prompts to 1 on JailbreakBench, from 304 of 313 to 2 on StrongREJECT, and from 198 of 240 to none on HarmBench, while GPQA Diamond and τ-bench capability scores were unaffected. The company said it contacted all impacted customers, recommended they review outputs from the period if their use depended on the model's refusals, and confirmed no other models, clusters, keys or customer credentials were affected.

read1 min views4 publishedNov 19, 2026
Prime Intellect says DeepSeek v4.1 Flash swap did not expose prompts
Image: source

Between November 8, 2026, at 01:24 UTC and November 10, 2026, at 06:35 UTC, an unauthorised party replaced a deployment behind Prime Intellect’s deepseek/deepseek-v4.1-flash endpoint with an abliterated copy, meaning a copy modified to reduce or remove refusal behavior. Requests served by the affected deployment during this period were answered by this model.

The model was replaced with publicly available abliterated weights of deepseek/deepseek-v4.1-flash. The chat template, tokenizer and serving configuration were not modified. We tested the modified weights and found that capability benchmarks (GPQA Diamond and τ-bench) were unaffected. Refusals fell from 98 of 100 prompts to 1 on JailbreakBench, from 304 of 313 to 2 on StrongREJECT, and from 198 of 240 to none on HarmBench. All customers impacted by the incident have been contacted. If their use of this model depended on the models refusals, we recommended to review their outputs from this period.

We have found no evidence that prompts or completions were recorded or left our deepseek/deepseek-v4.1-flash serving cluster, and no evidence of access beyond it. No other models on our platform were modified and no other clusters or infrastructure were reached. No keys or customer credentials were affected.

Our endpoints are set up with ZDR and do not log any requests. We recovered every version of the attacker's proxy, examined the disks and memory of every node involved, and hold process-level records of their sessions; none reads or transfers request content. Our investigation is ongoing, and we will update this notice if this assessment changes.

── more in #ai-safety 4 stories · sorted by recency
── more on @prime intellect 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/prime-intellect-says…] indexed:0 read:1min 2026-11-19 · —