AI SECURITY
30 items
30 posts
Prompt injection is role confusion, per an ICML 2026 paper: LLMs judge who is speaking by style, not role tags, so text that sounds like the system wins.
MCP Enterprise-Managed Authorization explained: zero-touch OAuth moves agent tool access into the identity, policy and audit systems enterprises already run.
Anthropic's open-source vulnerability harness shows AI security agents need reproducible exploit loops, separate verification agents, and patch receipts.
Anthropic's Claude containment writeup points to the next security layer for coding agents: deterministic capability ledgers, not another approval prompt.
The ChatGPT for Google Sheets exfiltration report shows spreadsheet agents need permission ledgers: action-scoped, logged, revocable, and visible permissions.
Before an AI agent gets tools, files, APIs, MCP servers, or deployment access, decide what it can read, write, call, log, and roll back.
Permissions, logs, and rollback for AI coding agents work best as one system. A run ledger and practical allow, ask, and deny policy for agents that edit code.
Prompt injection in agent apps is an architecture problem. Practical defenses: data boundaries, structured handoffs, tool guardrails, and approval gates.
AI security scanners find vulnerability candidates faster than teams can verify, disclose and patch them. Why triage is now the real bottleneck.
Prompt injection in CONTRIBUTING.md and AGENTS.md: AI coding agents read repo docs before opening PRs, so those files are now part of the security boundary.

Get Smarter About AI Dev
New tutorials, open-source projects, and deep dives on coding agents - delivered weekly.