/now
What I'm working on right now
Last updated:
Current focus
- Hardening rubica's AI platform ahead of the AU finserv pen-test — supply-chain gates, CodeQL coverage, signed commits, CycloneDX SBOMs on release.
- Native Anthropic regression work — fixing unsupported-content-format edge cases across our multi-provider message converter.
- Bounty-eligible CVE research across OSS AI infrastructure — MCP server implementations, LLM-app middleware, and agent SDK rendering paths. Following the huntr.com disclosure track; CVE-2026-44429 was the first one through.
- Upstreaming Model Context Protocol improvements — recent merged PRs on the TypeScript SDK (#1875, PrivateKeyJwtProvider custom claims) and Registry (#1145, #1149). Scoping the next one or two around auth flexibility and publisher tooling.
- Drafting the Lessons series — long-form essays from rubica war stories: token-budget pruner, MCP 64-char outage, cache_control TTL drift, supply-chain hardening. Each piece a production decision and the trade-offs behind it.
Recently shipped
- CVE-2026-44429 — disclosed and fixed a stored XSS in the MCP Registry catalogue UI.
- Token-budget pruner— pre-flight guard that physically removes MCP tools from the per-turn payload when overhead crosses 60 % of the model's context window.
- Signed-commits enforcement on main across all rubica engineering repos (SSH signing path, not GPG).
- CalVer release pipeline for rubica-content-hosting — independent of the rubica-chat stream.
Reading / following
- Model Context Protocol spec changes — particularly tool-name length and content-block schema drift between providers.
- Anthropic engineering posts on Claude Code subagents + skills — building out my own dev-team locally.
- APRA CPS 234 control mappings — incident response playbooks aligned to NIST 800-61 lifecycle.
Location
Istanbul, Türkiye. Remote for rubica (Hobart / Sydney).