Topic

#Llm Security

13 articles on Llm Security — news, releases, guides and analysis from the SourceFeed engine.

The frontier red-team playbook, three years on

Anthropic's 2023 biosecurity exercise hardened into industry law, and app teams keep copying the wrong half of it.

The exfiltration flaw sat unfixed for months, and the mitigation admins would reach for doesn't actually work.

Researchers replayed encrypted chain-of-thought into weaker sibling models and pulled live credentials out of public agent logs.

GLM-5.3's post-training gains are real enough that Z.ai delayed its own weights release.

Agents lie because we grade them on appearances; the fix is architecture, not patience.

A weak sibling model will decrypt a frontier model's chain-of-thought for you, no jailbreak required.

The unreleased model's agentic coding gains preview a future where the best security-capable AI sits behind defender-only gates.

Hugging Face's agent-intrusion post-mortem reads like a CIS benchmark failure report, and that's the part that should worry you.

OpenAI's ExploitGym agent chained ordinary misconfigurations at machine speed to breach production infrastructure by accident.

Run dozens of adversarial probes against your LLM endpoint on every PR and fail the build on regressions.

A record open-weight release, Xi's open-source pitch, and a breach where the only usable model was Chinese.

LLMs assign authority based on how text is written, not where it comes from, making role tags a leaky abstraction.

NVIDIA's open-source tool introduces capability governance to protect agentic workflows from prompt injection, tool poisoning, and malicious skills.