14 Agentic AI Security Projects Every AI Security Engineer Should Build
Full-Presence7590 · reddit · 2026-11-04
An AI security engineer lays out 14 hands-on projects for securing autonomous agents, each with a rationale:
- Crescendo attack fuzzer: a multi-turn stateful harness that grooms context windows to bypass commercial system prompts — single-shot injections are dead.
- Sandboxed MCP proxy: JSON-RPC middleware validating MCP tool calls via Pydantic to block confused-deputy escalation; tool poisoning is "the SQL injection of the agentic era."
- Inter-agent sanitization bus to strip hidden instructions between agents in swarms; PII UUID-swapping gateway so raw PII never touches frontier APIs; RAG poison detector cross-referencing chunk hashes.
- CFG output caging at token-decoding level ("system prompts are suggestions, grammar constraints are physics"); time-limited scoped OAuth JWTs for delegated agents.
- Also: PyRIT CI/CD gates blocking merges above 1% jailbreak rate, semantic rate-limiting against weight-distillation bots, pixel-level auditing for computer-use agents, vector-DB sleeper-agent scrubbing, hash-chained audit ledgers, and SBOM scanning of vibe-coded output.
More from coding & agent
- Shopify CEO hails Tomo agent buying via Shop Pay over iMessage — Jackyhuang · 2026-11-04
- Apple opens Siri to all apps: chaining, screen awareness, dev beta — SumitGup · 2026-11-04
- Stripe hires engineers to crack hard problems in agentic commerce — jeff_weinstein · 2026-11-04
- xAI engineer yacineMTB: agents work autonomously while I sleep, I wake up to breakthroughs — yacineMTB · 2026-11-04
- 8 Coding Agents on One RTX PRO 6000: Measured Throughput for Local Qwen Deployment — AIFlow_ML · 2026-11-04
- A start-to-finish workflow for AI dialogue scenes, from script to 1080p — DavidmComfort · 2026-11-04