AI application security
Prompt injection, retrieval trust, sensitive data exposure, and the boundaries between model outputs and application behavior.
UNTRUSTED INPUT → CONTROLLED EXECUTIONAI SAFETY & SECURITY
AI moves fast.
Trust takes
evidence.
Large Metric Model is building a security practice for the age of capable AI. We focus on the applications, permissions, and decisions that turn model behavior into real-world risk.
INDEPENDENT PERSPECTIVE. MEASURABLE QUESTIONS.
TRUST BOUNDARY / 001
Define. Test. Verify.Concept study ↗
THE LMM PRINCIPLE
Measure what AI can do.
Test where it breaks.
01 / OUR FOCUS
Security at the point
where AI takes action.
Our practice is taking shape around three connected areas. Each starts with a specific failure mode and ends with evidence that a defense works.
↗
Agent safety evaluations
Tool permissions, approval steps, task scope, and the visibility needed to understand what an autonomous system actually did.
CAPABILITY → ACCOUNTABLE ACTIONDefensive engineering
Threat modeling, secure code review, finding validation, and regression tests that connect a security concern to a practical fix.
REPRODUCE → PATCH → RETEST02 / RESEARCH AGENDA
Make the question
testable.
We are developing our first studies. Explore the questions, proposed methods, and evidence we intend to publish.
PLANNED STUDY / 01
When documents try to give orders.
Can an untrusted document redirect an agent or expose a synthetic secret?
- Test design
- Use a small document corpus with seeded adversarial instructions and fake credentials. Hold the task constant while changing input trust controls.
- What we measure
- Unauthorized tool calls, synthetic secret exposure, and legitimate task completion.
- Evidence to publish
- Versioned fixtures, model and tool configuration, complete action traces, and results before and after a mitigation.
Design stage · results will follow completed testing
03 / OUR APPROACH
Permission first.
Evidence throughout.
Our operating principles for security research and defensive work.
Explore the OWASP LLM risk taxonomy ↗Define the boundary.
Document the systems, actions, and permissions in scope. Begin with company-owned test environments and synthetic data.
Make failures reproducible.
Record the model, version, inputs, tool configuration, and action trace. Measure both security failures and useful task completion.
Test the defense.
Review findings, apply a focused mitigation, and rerun the original case. Explain remaining uncertainty alongside the result.
Publish with care.
Plan disclosure with affected maintainers. Share reproducible lessons and redacted evidence after findings are validated.
ABOUT LMM
Following the boom.
Examining the risk.
LMM stands for Large Metric Model, the public brand of LMM AI LLC. We are an early-stage AI safety and security firm developing a practice around rigorous evaluations and practical defensive engineering.
We follow emerging AI capabilities with a simple question. What evidence would make this system safer to use?
Follow the work at @LMM ↗LET’S COMPARE NOTES
Building with AI?
Start with the hard questions.
Connect with LMM about AI security, evaluation methods, or a research collaboration.
Connect with @LMM ↗Public updates and initial contact on X.