Cyera flags trust gaps in LLM tools and coding agents

Cyera Security Research Team Lead Ofek Itach traces his path into research to defensive work on network protocols, where he learned to question the assumptions developers build into systems. Cyera’s security research team studies AI security, data security, and vulnerabilities, with findings feeding into product coverage, proofs of concept, and guidance on emerging areas of risk.

The team uses AI as a force multiplier rather than a replacement for researchers. It scans code with simple filters, applies AI to reduce false positives, and sends promising leads for human review. AI has reduced the time needed to find bugs, but the decisive work still comes from reading code directly.

Cyera’s most significant finding to date was a critical vulnerability in Ollama, an open-source tool for running LLMs in-house. The unauthenticated remote memory read could let an attacker without credentials access raw server memory, exposing conversations, prompts, passwords, tokens, and other secrets.

Itach sees coding agents as a major unresolved trust problem because they are often granted broad machine permissions to read, write, and run tasks. Cyera is focusing on the point where that trust can fail, especially when the same environment contains credentials, private source code, customer data, and secrets.