claude-code-hooks: Runtime Least Privilege for AI Coding Agents
Selected for Arsenal in London. A demonstration of the open-source claude-code-hooks collection and runtime controls for AI coding agents.
Talks and workshops on agentic AI for security engineering: autonomous AppSec agents, MCP, and the guardrails that make AI coding agents safe.
Selected for Arsenal in London. A demonstration of the open-source claude-code-hooks collection and runtime controls for AI coding agents.
Five least-privilege patterns for agents and MCP servers, with a public companion repository to reproduce the demos. Recorded for the virtual conference; scheduled for 15:00 to 15:30 IST.
Delivered remotely for the Auckland conference. Practical hooks for controlling what AI coding agents can read, change and execute.
Recorded two weeks after OWASP re-ranked the LLM Top 10 for 2026, this session maps deterministic hooks to the new list and runs live-fire demos against the open-source 12-hook set: secret reads denied, curl piped to shell blocked, sudo escalated to a human ask, a failing test protected from deletion. Recording the demos surfaced two real bypasses in the secrets guard; the failed takes shipped as patches with regression tests.
Enforcing least privilege on autonomous developers. AI coding agents execute tool calls with your credentials. Every one of them is attack surface. How hooks intercept every tool call before it runs, mapped to the OWASP LLM Top 10: blocking dangerous commands, protecting secrets, and enforcing policy in about 100 lines of Node. Based on the open-source claude-code-hooks collection.
Conference talks, workshops, and panels on agentic AI for security engineering.
karan@karanbansal.in