Skip to content
View sushant-me's full-sized avatar

Block or report sushant-me

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
sushant-me/README.md
Sushant Poudel — AI security engineer. I find where agent tool boundaries leak, then close them.

AI security engineer — I find where agent tool boundaries leak, then close them.

Bhaktapur, Nepal · open to security-engineering and AI-security work, remote or on-site

sushantpoudel2028.com.np · Writing · Publications · Press · LinkedIn · sushant.poudel2028@gmail.com

What I do, and how to work with me →


Receipts

Chosen because each one can be checked from the link next to it.

  1. A merged security fix in google/go-github. #4556 stopped a release-asset upload from being sent to a host the client never configured. A maintainer generalised the rule the next day in #4564 — "send credentials only to configured origins" — and merged #4579, which puts Sushant Poudel in AUTHORS on master and in the v93.0.0 release notes.

  2. A bug filed, confirmed by the maintainer, and fixed the same night. google/osv-scanner compiles nameIsRegex overrides by string concatenation, so a top-level alternation escapes the documented ^pattern$ anchors and silently drops packages from a vulnerability scan: #3120. The maintainer replied "what you have pointed out is true" and merged his own fix, #3121, hours later.

  3. Memory-safety findings with standalone reproducers. Five open issues in google/s2geometry — a NULL dereference (#674), an out-of-bounds read (#676), a heap-buffer-overflow (#678), a 16 GiB allocation from a 117-byte input (#677) and a 2.4 GiB allocation from a 28-byte input (#679) — plus two hardening PRs (#681, #682). A heap-buffer-overflow write in pnggroup/libpng (#935), reproduced under AddressSanitizer on 1.6.60 and 1.8.0; stated limit: caller-triggered, not reachable from crafted PNG bytes. And an OSS-Fuzz harness in google/libphonenumber that was asserting 0.00% coverage of the code it targeted — the corrected harness measures 93.05% locally (#4079, open).

  4. Four published security advisories against my own tools — one high, three medium, each found, fixed and written up: policygate — a newline stepped around a deny glob, an authorisation bypass in the gate itself; mcp-nameguard — an unbounded server reply let a hostile MCP server exhaust the scanner inspecting it; agentbound — the masker blanked one-line strings and hid real findings; mcpaudit — invisible-character ranges covered 16 of 256 variation selectors.

  5. HackingHub — Elite, 21,785 XP, 61% of the platform conquered, 1 gold / 4 silver / 3 bronze medals, #7 on the all-time global leaderboard. Profile · leaderboard API

  6. NullOrigin CTF 2026 — grand finalist, rank #2. Certificate NO-2026-GF-229CF4 verifies at creds.cyberhx.com/verify; the event ran a 12-hour qualifier and a 12-hour final, with 50 finalist teams selected from 3,500+.

  7. A paper with a DOI — Edge-Native Semantic Firewall for Autonomous LLM Agents: A Structured Chain-of-Thought Verification Framework, Proceedings of ICICSET 2026, published 2 October 2026: 10.65091/icicset.v3i1.75.

Other merged upstream work — breadth, if you want it
  • open-telemetry/opentelemetry-go — two merged (OTLP log-exporter env-var case handling, #9230; module version reporting, #9232), one open (#9231).
  • google/adk-python — the framework now refuses its reserved set_model_response tool name (commit cc5d516, 2 Oct 2026). google/adk-go — two merged (#1679, #1711), one open (#1606).
  • trailofbits/skills — two merged (#346, #347); trailofbits/it-depends — one merged (#254).
  • sisaku-security/sisakulint — a docs fix merged (#655); three issues from an audit of its AI-action rules remain open (#646, #647, #648) after the maintainer reproduced the finding independently.
  • elementalsouls/Claude-BugHunter — routing fix merged (#81): 22 of its 58 hunt-* skills had no mapping and could never be selected.

What I do

Agent and MCP security. Detectors, audits and fail-closed gates for the boundary where an agent's tool declarations become executed actions — agentbound, mcpaudit, policygate, trajectorycheck, and the labelled corpus that keeps their precision numbers honest.

Memory safety and fuzzing. ASan/UBSan triage, harness engineering, and upstream reports with reproducers — libpng, s2geometry, libphonenumber's dead OSS-Fuzz harness, OpenTelemetry-Go.

Full-stack and offline-first builds. Next.js/TypeScript and Flutter products for infrastructure that is not reliable: two years as an AI engineer at Atmos SoftTech (Mar 2024 – Mar 2026) building production ML, agent pipelines and offline inference. meshdoc is the current one, in Rust.

Background: Computer Engineering student at Nepal Engineering College, Bhaktapur.

Selected work

  • agentbound · Python, ships as a GitHub Action · static detection of agent tool-boundary bugs across Python, TypeScript, JavaScript, Go, Java and Actions workflows. Precision measured, not claimed: 0.79 → 1.00 across a 14-framework survey, and 0.94 once two more frameworks were added in v0.1.16 — every false positive, the rule that removed it and the four mechanisms behind them are in SURVEY.md. Its advisory fixed a masker that was hiding a real finding.
  • mcpaudit · Python · audits an MCP server's tool declarations before you connect, then pins them in a lock file — a description that changes after approval is reported as the tool-poisoning shape it is.
  • policygate · Python · a fail-closed gate for agent tool calls: the model may deny or escalate but cannot authorise, a failing evaluator falls through to refusal, and every decision lands in a hash-chained audit log. Its high-severity advisory was an authorisation bypass I found in my own gate.
  • trajectorycheck · Python · grades whole agent trajectories — tool calls, arguments, side effects, injected instructions — with a deliberately broken agent as a control.
  • tool-boundary-corpus · Python · 24 labelled tool-boundary cases and a detector-agnostic harness, so a precision number can be re-measured instead of trusted.
  • Edge-Native Semantic Firewall · Python · the paper's artifact: 600-scenario corpus, raw model outputs, and a result that contradicted the hypothesis — JSON-constrained output without a reasoning field approved 46.2% of policy-blocked proposals, against 17.2% free-form.
  • meshdoc · Rust · offline-first, end-to-end-encrypted collaboration for teams without reliable connectivity.
  • reputation · Python · every claim I make about my own work, its public source, and the weekly check that turns red when one stops being true.

Six tools on one page, with install lines and Action snippets: sushant-me.github.io/tooling.

Publications and writing

Work with me

I take on work where the deliverable can be checked: MCP server and agent-tool-boundary audits, threat models of an agent's action surface, memory-safety and fuzzing reviews of the C/C++ underneath, and full-stack builds where connectivity is not a given.

Hire page → · CV (PDF) → · sushant.poudel2028@gmail.com · LinkedIn · Press and author page

Every claim above links to the thing that proves it, and where a claim has a limit the limit is in the sentence. The subset that is machine-checked every week lives in reputation. Work that cannot be linked — private bug-bounty reports, unpublished audits — is deliberately not listed here.

Pinned Loading

  1. agentbound agentbound Public

    Static detector for AI-agent tool-boundary bugs (tool shadowing, reserved-name collisions, fail-open confirmation gates, unauthenticated agent CI dispatch)

    Python

  2. mcp-nameguard mcp-nameguard Public

    Check an MCP server's tool names against the names agent frameworks reserve for their own tools (tool shadowing) · GitHub Action

    Python

  3. trajectorycheck trajectorycheck Public

    Trajectory-level evaluation for AI agents — grades tool calls, arguments, and side effects, not just the final answer

    Python

  4. tool-boundary-corpus tool-boundary-corpus Public

    A labelled corpus of agent tool-boundary cases, and a detector-agnostic harness that scores precision and recall against them · GitHub Action

    Python

  5. reputation reputation Public

    Every claim I make about my own work, and the script that re-checks all of them against their public sources

    Python

  6. Edge-Native_Semantic_Firewall_ Edge-Native_Semantic_Firewall_ Public

    Edge-native semantic firewall for autonomous LLM agents: a 600-scenario evaluation of structured Chain-of-Thought policy verification on a 3.8B model running locally within a 4.2 GiB VRAM budget.

    Python