Executive Summary
Researchers reported risks involving an OpenAI Codex repository and backdoored model weights, including potential malicious package publication and triggered agent commands.[1] Organizations running coding agents, LLMs, or AI supply chains should prioritize provenance, integrity, and adversarial testing.
What Happened
Noteworthy stories that might have slipped under the radar: Siemens details security advisory policy, OpenAI Codex backdoor risk, Microsoft warns of model weight backdoors. The post In Other News: New CISA Director Nominee, Connected Car Hacking, Baker Tilly Data Breach appeared first on SecurityWeek .
Affected Products
- OpenAI Codex
- Codex npm package
- Nemotron-Terminal-8B
Vulnerability / Attack Technique
Prompt injection and model-weight backdooring
Technical Mechanism
The reported Codex attack began when an issue-labeling workflow passed issue content to a Codex agent; the chain could allow sandbox code execution, access to tokens in a GitHub runner process, workflow-cache poisoning, and a malicious npm release.[1] Microsoft researchers reported that modifying fewer than 10 dormant model neurons could create a conditional backdoor that behaves normally until a trigger, such as a date or error message, causes the agent to emit an attacker command.[1]
AI Security Category
LLM03:2025 Supply Chain
Part of the AI Supply Chain and AI SBOM topic.
Relevant Frameworks
RealGround Analysis
AI agents can convert untrusted repository content into execution paths and may expose build credentials or trusted publication mechanisms when granted excessive permissions. Backdoored model weights create a supply-chain risk that may evade ordinary behavioral checks because malicious behavior can remain dormant until a specific trigger appears; teams should compare deployed weights with trusted originals and assess agent actions under adversarial inputs.
Recommended Controls
- Pin and cryptographically verify model files, package artifacts, dependencies, and workflow actions against trusted provenance records.
- Run issue and repository content through strict prompt-injection defenses before providing it to coding agents.
- Use least-privilege, short-lived tokens and isolate agent sandboxes from CI runner memory, package registries, and production credentials.
- Require human approval and independent verification before publishing packages or modifying workflow caches.
- Red-team agents and model checkpoints with trigger-based backdoor tests, and compare deployed model weights against known-good originals.
Related RealGround Service
AI Agent SBOM and LLM Mapping · or start with the free AI Agent Security Scanner.
Sources
securityweek.com: https://www.securityweek.com/in-other-news-new-cisa-director-nominee-connected-car-hacking-baker-tilly-data-breach/
Summary and analysis written by RealGround from the source above; last updated 2026-11-14.
