Muhammad Hasnain

Independent researcher, AI-agent security · founder, Bin Abdullah LLC (bryteai.studio), NVIDIA Inception member

I study how AI agents get hijacked by the content they read: where prompt-injection detectors fail, and where the trust boundary should sit.

GitHub PyPI Hugging Face Email

Research

toolwall: a firewall between an AI agent and its tools

Open source, Apache-2.0 · Python · v0.4.0, October 2026

Every tool output (web page, email, file, MCP result) is checked before the model reads it: policies, rules, and a small classifier that reads long outputs in windows. One decorator, or a proxy in front of any MCP server.

82.2 %
of 500 human-written LLMail-Inject attacks blocked, 0 of 703 benign inputs blocked. Best other open guard: 44.9 %.
1.12 %
of 9,160 clean SEP rows blocked by mistake, measured for eight guards (0.09 % to 2.49 %).
fails
on LongPIBench long documents, like every guard tested. Stated in the report, not hidden.
0 / 19,173
benign texts changed by the v0.4 trust-boundary filter; matches the SecPO chat template on 20,000 random strings.

Protocol: test sets frozen before training, every guard run once, per-row decisions published. Report · Trust-boundary design note

prompt injectionLLM agentsMCPevaluation

Reports and artifacts

Product

Landlord Tax Lab

UK landlord tax calculators, 2026/27 rules

Rental-income, capital-gains, stamp-duty and Making Tax Digital calculators. 402 clicks and 69,651 impressions from Google Search in the 90 days to 16 October 2026.

Learning