Agent authority
Identity, delegated permissions, tool contracts, approvals, revocation and non-human accountability.
Open research / versioned sources
Standards are inputs. Evidence is the output.
Delx Security conducts defensive AI-agent research by versioning primary sources, bounding safe experiments and publishing only reproducible, decision-useful evidence.
Research agenda
Context integrity
Prompt injection, memory poisoning, retrieval trust, cross-agent influence and durable contamination.
Safe execution
Sandbox boundaries, egress, secrets, supply chain, side-effect controls and emergency interruption.
Assurance evidence
Runtime proof, reproducibility, control drift, verification quality and decision-useful reporting.
How the research becomes evidence
Defensive AI-agent research is a bounded evidence loop, not a framework checklist. Each question records its sources, scope, smallest reproducible observation and remaining uncertainty.
Version primary sources
State the question, why it matters and which primary sources and versions establish the baseline.
Bound the environment
Work only in an authorized or synthetic environment with explicit controls, timing and stop conditions.
Reproduce the observation
Capture the smallest reproducible observation, its impact and confidence without collecting unnecessary sensitive data.
Publish limits and retest
Separate fact, inference and unknown; state the defensive implication, residual risk and how a change can be verified.
Primary sources
NIST Cybersecurity Framework 2.0Official source ↗NIST AI Risk Management FrameworkOfficial source ↗OWASP Top 10 for Agentic Applications 2026Official source ↗OWASP Application Security Verification Standard 5.0Official source ↗MITRE ATLASOfficial source ↗CISA Secure by DesignOfficial source ↗NIST Secure Software Development FrameworkOfficial source ↗SLSA 1.2 supply-chain specificationOfficial source ↗OWASP AIBOM GeneratorOfficial source ↗MCP 2026-07-28 Tools specificationOfficial source ↗RFC 9116 — security.txtOfficial source ↗
Publication standard
Research notes distinguish verified facts, inferences, experiments and open questions. Versions and retrieval dates stay attached to changing frameworks.
We do not publish exploit details that create disproportionate harm, client evidence, private infrastructure or claims that cannot be independently supported.