5 min read
Research LLMs with large context windows can be reliably jailbroken by embedding hundreds of fictitious dialogues before the target request. The attack scales with context length.
All Articles
4 min read
Vulnerabilities CVE-2026-31204: an SSRF vulnerability in Ollama allows local-network attackers to read arbitrary files and reach internal services via the model pull endpoint. All versions affected.
5 min read
Research Backdoor behaviours introduced into LLMs during fine-tuning can persist through subsequent safety alignment including RLHF and adversarial training, surviving standard red-teaming.
7 min read
Defensive Techniques A practical framework for implementing prompt injection detection at the API gateway layer: input sanitisation, context isolation, output filtering, and anomaly detection.
3 min read
Vulnerabilities A critical deserialization vulnerability (CVSS 9.8) in a widely-deployed ML model serving framework allows unauthenticated RCE via crafted model files. Active exploitation confirmed.