large language models Security Incident

Researchers demonstrated that as few as 250 poisoned documents can create a backdoor vulnerability in large language models, irrespective of model size or training data volume. This data poisoning technique, which can induce denial-of-service or potentially facilitate data exfiltration, challenges prior assumptions about the required scale of malicious training data during pretraining.

What Happened

Researchers demonstrated that as few as 250 poisoned documents can create a backdoor vulnerability in large language models, irrespective of model size or training data volume. This data poisoning technique, which can induce denial-of-service or potentially facilitate data exfiltration, challenges prior assumptions about the required scale of malicious training data during pretraining.

Why This Matters

Publisher reporting describes a security event affecting LLMs. BugSkan could not yet bind a CVE or affected version, so treat the source details as the current record.

Recommended Action

Confirm whether LLMs is present in your environment and review vendor guidance for this report. Apply available patches or mitigations if your deployment matches the described conditions.

Exposure

Exposure reason: This incident does not currently match a technology in My Interests.

Exploitation status: UNKNOWN

Primary entities:

Timeline

• Incident first seen
  Oct 09, 2025 05:30

  BugSkan first recorded this incident.

• A small number of samples can poison LLMs of any size - Anthropic
  Oct 09, 2025 05:30

  anthropic.com · Vulnerability

Sources

Related Incidents

Other BugSkan incidents that share identifiers, products, or vendors with this report.

My Interests Match