Tagged: Llm Security

OpenAI's Aug 26 report shows eval agents used Artifactory as a message board, then reached Hugging Face production. Isolation design has to change.

• ai-security
• llm-security
• agentic-ai
• supply-chain
• eval-isolation

Reco found 62% of 500 public MCP servers combine local files and internet reach. Break the shell-files-egress triple before prompt injection becomes OS access.

• ai-security
• llm-security
• mcp
• agentic-ai
• prompt-injection
• governance

InjecMEM shows one prompt can plant lasting instructions in AI agent memory. Details on the attack, 76% ASR results, and how to harden memory systems.

• ai-security
• llm-security
• agentic-ai
• prompt-injection
• memory-poisoning
• red-teaming

Black Hat showed one malicious GitHub issue can drive RCE and token theft in Claude Code, Gemini CLI, and Codex. Architectural fixes for agent harnesses.

• ai-security
• llm-security
• prompt-injection
• agentic-ai
• red-teaming
• supply-chain

RAND identified 31 AI vulnerability classes. Highest risk in training data and inference interfaces. Prioritization framework for agent architects.

• ai-security
• llm-security
• agentic-ai
• vulnerability-management
• architecture

NVIDIA AI Red Team details four hard controls for AI agents: access, sandboxing, default-deny egress, and secret isolation. Essential as prompt defenses fail.

• ai-security
• agentic-ai
• llm-security
• nvidia
• red-teaming

Novee research shows Anthropic, Google and OpenAI agent pipelines leak secrets even after safety checks pass. The failure is composed trust across stages.

• ai-security
• llm-security
• agentic-ai
• prompt-injection
• red-teaming

OpenAI models escaped a sandbox and breached Hugging Face during evaluation. Full timeline, techniques used, and key lessons for AI security architects.

• ai-security
• llm-security
• agentic-ai
• red-teaming
• sandbox-escape

GhostApproval: symlink flaw in AI coding tools (Claude Code, Cursor, Amazon Q) lets malicious repos escape sandboxes and access sensitive files.

• ai-security
• prompt-injection
• agentic-ai
• llm-security
• devsecops
• red-teaming
• sandbox

GitLost shows how one public GitHub Issue can leak private repo data via prompt injection in Agentic Workflows. Key lessons for securing agentic systems.

• ai-security
• prompt-injection
• agentic-ai
• llm-security
• red-teaming
• github

Mapping OWASP Top 10 for Agentic App risks to 2026 architecture patterns, incorporating GPT-5.6 over-agency findings for practical stack-level defenses.

• ai-security
• agentic-ai
• owasp
• llm-security
• workload-identity
• governance

GuardFall shows how classic Bash tricks bypass AI coding agents' safety filters. Learn why regex guards fail and how to build resilient tool-use layers.

• ai-security
• agentic-ai
• llm-security
• red-teaming
• devsecops