Tagged: Prompt Injection

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

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

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

Dual-firewall and adaptive out-of-band defense patterns for securing agentic and RAG systems against prompt injection.

• ai-security
• prompt-injection
• agentic-ai
• rag-security
• llm-agents
• multi-agent-systems