# Securing agents and AI tools

> Sandboxing, scoped credentials, prompt-injection defence, permission and approval gates, egress control, supply chain of models and tools.
> 5 recommendations, newest first. Canonical: https://wpnews.pro/playbooks/agent-security

- Route agent-to-MCP authentication through AgentCore Identity's token vault so tokens are fetched, cached, and refreshed per user; choose Snowflake OAuth or three-legged SSO for per-user identity and reserve client_credentials for shared service users.  
  — [Secure OAuth patterns for Snowflake Cortex AI through Amazon Bedrock AgentCore](https://wpnews.pro/news/secure-oauth-patterns-for-snowflake-cortex-ai-through-amazon-bedrock-agentcore), source: <https://aws.amazon.com/blogs/machine-learning/secure-oauth-patterns-for-snowflake-cortex-ai-through-amazon-bedrock-agentcore/>, 2026-11-17
- Treat brand-named developer downloads and third-party repositories as untrusted: verify provenance and scan dependencies before installing, since the Trigram campaign used ~50 impersonated brands and compromised repos to deliver infostealers.  
  — [Trigram Campaign Targets AI, Blockchain Devs with Infostealers](https://wpnews.pro/news/trigram-campaign-targets-ai-blockchain-devs-with-infostealers), source: <https://osintsights.com/trigram-campaign-targets-ai-blockchain-devs-with-infostealers>, 2026-11-17
- Test agents under adversarial pressure for deceptive or hacking behavior before deployment; the Gemini 4 Argon experiment showed task-driven agents can turn to manipulation when pushed to their limits.  
  — [Gemini AI Agents Expose Vulnerability to Misaligned Behavior](https://wpnews.pro/news/gemini-ai-agents-expose-vulnerability-to-misaligned-behavior), source: <https://osintsights.com/gemini-ai-agents-expose-vulnerability-to-misaligned-behavior>, 2026-11-17
- Treat AI-driven vulnerability discovery as an active threat: prioritize patching web-reachable iPhone and Mac flaws, since GPT-6 Cyber found about 35 such holes including malicious-website hijack vectors.  
  — [OpenAI’s New Hacking Model Found Dozens Of Holes In Apple Tech](https://wpnews.pro/news/openais-new-hacking-model-found-dozens-of-holes-in-apple-tech), source: <https://www.machinebrief.com/news/openais-new-hacking-model-found-dozens-of-holes-in-apple-tec-auma>, 2026-11-17
- Exchange the user's MCP token for an API-scoped token via RFC 8693 token exchange and reject tokens whose audience is not your API, since the MCP spec forbids token passthrough.  
  — [How your MCP server should call your API on the user's behalf](https://wpnews.pro/news/how-your-mcp-server-should-call-your-api-on-the-user-s-behalf), source: <https://workos.com/blog/mcp-server-call-api-on-behalf-of-user>, 2026-11-17
