Software engineer. Agent tooling, eval harnesses, vulnerability research.

• beardthelion.dev
• X

Reported vulnerabilities:

In triage (details withheld until the vendor publishes):

Detection work:

• cve-variant-sweep: Validated Semgrep rules from recent guard-adding CVE fixes, for variant hunting. 75 verified rules, all pushed with test evidence. One generalized variant flagged a live sibling instance in Pillow that the advisory fix missed.
• RedPen Eval: Eval harness that grades AI code reviewers against real CVE-derived vulnerability cases. 48 seeded cases, ground truth from the sweep rule corpus.
• ssrf-guard-audit: Guard-completeness audit for SSRF defenses: classifies how each caller-controlled URL fetch site is incomplete, not just whether a sink exists.
• mcp-ssrf-audit: The same guard-completeness audit applied to MCP servers, with Semgrep rules, CLI, and labeled corpus.
• path-guard-audit: Static audit of incomplete filesystem and archive path guards in Python and TypeScript.
• egress-plane: Static egress scanner across Rust, Python, JavaScript, TypeScript, and Go, with blind spots declared and completeness reported in its exit code.
• dyetrace (private): Black-box dataflow tracer: injects a unique tracer into every observed input of a recorded session, re-drives read surfaces, and flags tracers resurfacing where they should never reach.

• signet: Self-sovereign agent state: DID-rooted, end-to-end-encrypted, portable memory, config, transcripts, and permission grants as ciphertext the store can never read. Crypto-blind sync store, MCP adapter, conformance checker.
• memval: Eval harness: an agent with no memory versus encrypted memory backends, scored three-way with red-check controls.
• hermes-skill-distillation: Real-world agent trajectories for Hermes 4: task battery, judge, SFT export, Atropos RL.
• subprompt: Hermes plugin that resolves {{fuzzy description}} markers before the model reads the message.
• openbeard: Gemini CLI fork: OpenAI-compatible, MiMo reasoning, tool calling.
• hermes-sendblue-plugin: Sendblue iMessage/SMS/RCS for Hermes Agent.

Maintainer on the Twigpine org (formerly Gitlawb):

• twigpine/node: Path-scoped visibility on git and IPFS surfaces; withheld IPFS subtrees no longer leak through tree objects; peer rows bound to DID so only the keyholder can repoint them; peer-supplied repo slugs cannot escape repos_dir.
• twigpine/openclaude: Per-turn simple-vs-strong auto-routing; WebFetch SSRF-guard regression test; MCP skills from skill:// resources.
• twigpine/zero: Workspace-trust gate for hooks, plugins, and MCP; normalize launcher names before the sandbox command-prefix denylist.
• twigpine/openrecall: Zero-knowledge encrypted memory for AI agents. Hash namespace into slug so tenants cannot collide.

Upstream contributions:

• NousResearch/hermes-agent: 89 authored PRs, including opencode-go routing for qwen3.7-max via anthropic_messages.
• CopilotKit/OpenBot: MCP tokens only for their own server at their own address; refuse a credential written into the rest of the address; Docker socket on loopback, not every host address.
• ApeWorX/ape: Resolve overloaded methods by argument type, not just count.
• vercel-labs/fx: Write MCP config durably instead of truncating it in place; keep the MCP session id alive until in-flight requests drain.
• laude-institute/headlong: API keys off the docker command line; a sibling directory is not a mounted one.
• sooperset/mcp-atlassian: Convert markdown in Jira text-area custom fields to ADF.
• OpenListTeam/OpenList: Stop the upload retry loop when the context is canceled.

Rust, Python, JavaScript, TypeScript, Java with Spring and microservices, SQL.