ABOUT
Independent evaluation.
Accountable claims.
MurderBench is an AI agent evaluation initiative led by Linh Ngo and operated by XP.COM, LLC dba Xenocom.
Why this exists
Agents act through tools and safeguards. We want operators to know whether those safeguards enforce the constraint they care about, and whether a repair preserves the useful task.
We are building a focused engineering evaluation practice. We do not represent ourselves as a university, accredited research lab or a government-supported program.
Who does the work
Linh Ngo initiated and directs the initiative; the human headcount is one. He sets the research goals, priorities, sandbox boundaries and budget, challenges interpretations and requires novelty audits and publication of negative findings. AI coding agents drafted designs and source audits, wrote the implementation and analysis code, ran authorized experiments, and produced reports, figures and this website. Other AI agents supplied preparation critiques. These contributions do not establish unaided human implementation, independent human annotation or external peer review. Read the evidence and attribution overview.
No independent scientific reviewer, lab partnership or funding award was verified at the last outreach check on October 8, 2026. We are seeking critical review and will disclose material relationships. This October 15 evidence update does not add a model result or constitute a new inbox check.
What earns publication
A replayable result with clear denominators and uncertainty—or a documented reason to reject an idea. We preserve failed controls, incomplete collections and negative findings.
Read the evidence record →Linh Ngo on LinkedIn ↗Linh Ngo on X ↗Work with us
Review a concrete design, help reproduce a result, offer model access, or bring an agent safeguard that needs a bounded evaluation.
Contact ↗