Using MCP or LLMs Without Guardrails Is Irresponsible: Data Egress, Prompt Injection, Tool Logging
shensi · x · 2026-11-20
shensi argues that deploying MCP servers or LLMs without AI guardrails is irresponsible, outlining three core practices:
- Sensitive data egress control — catch sensitive data before it leaves the organization
- Prompt injection defense — block injections hidden in documents and tickets
- Full audit logging — log every tool call
In a follow-up, they note a second pain point: sharing AI skills across companies is still built for engineers. Skills typically live in GitHub with clunky processes that lock out anyone outside R&D, making them hard to adopt company-wide.
More from Safety
- Superforecasters put 33% odds on a week-long outage hitting 1M+ people in an OECD country by 2028 — NathanpmYoung · 2026-11-20
- Sentinel brief 'Consider Cyberarmageddon' extrapolates cyberattack growth to societal failure — NathanpmYoung · 2026-11-20
- NY senator moves to keep Waymo off streets; critics say delays will indirectly kill people — zacharynado · 2026-11-20
- About 100 protesters gather outside Anthropic's SF HQ to 'stop the AI race' — nmasc_ · 2026-11-20
- NVIDIA and ANSI to host physical AI safety session at GTC DC on Dec 2 — liu_mingyu · 2026-11-20
- Anthropic Safety Researcher Spars Over Wellbeing-Enhancement Stance — repligate · 2026-11-20