6 HIGH · 1 INFO · THREAT ORANGE · 7 items · Generated in 406s
Improper handling of image metadata can lead to unexpected rendering and distortion of input content, potentially affecting data integrity.
vLLM versions contain multiple ReDoS vulnerabilities, allowing attackers to trigger denial of service.
vLLM versions are vulnerable to crashes and resource exhaustion due to missing sparse tensor validation.
LangChain framework has a vulnerability that allows disclosure of files outside the intended boundary when receiving path values from an untrusted source.
Authenticated users can access unauthorized Ollama backends.
An unauthenticated remote information disclosure vulnerability allows an attacker to read and exfiltrate server's heap memory.
A challenge to hack an AI assistant was run, with 6,000 attempts made but no secrets leaked. The underlying model was Opus 4.6.