Attack-path reconstruction
Maintaining a current map of how an ordinary foothold reaches valuable systems, updated as identities and configurations change.
Research program
AI-native defensive security
AI makes some attacks cheaper and gives defenders more ways to watch systems continuously. This program asks which neglected defensive jobs machines can now do, and which decisions still need a person.
Which defensive work should machines carry, and which decisions must remain with people?
The division of labour
Claim: recommendation
Machines should increasingly do the remembering, the monitoring, the correlation, the testing, the evidence gathering, and the routine enforcement. People should make the decisions that carry consequence: isolate a system, disclose an incident, notify customers, invoke counsel, halt production, or accept a risk.
Models will make mistakes. Some mistakes can be caught on the next pass. An automated disclosure sent in error cannot be recalled, so a person keeps that decision.
Work that becomes possible at continuous scale
Claim: hypothesis
These jobs were sampled, scheduled, or skipped because nobody could watch them all the time. Machines may now run them continuously. The work still has to show that its error rate is acceptable.
Incident timeline construction
Assembling a defensible account of what happened from logs across systems, with each statement tied to its source record.
Control verification
Repeatedly exercising controls against the behaviours they claim to stop, rather than testing once a year.
Secrets and credential discovery
Continuous scanning of repositories, images, logs, and third-party surfaces for exposed material, with automated revocation paths.
Policy to evidence mapping
Connecting a written requirement to the specific operational query that demonstrates it, and flagging requirements with no source.
Adversarial and phishing simulation
Generating varied, current scenarios rather than reusing a template, then measuring behaviour rather than completion.
AI theatre is the obvious failure
Claim: inference
Every criticism this program makes of security products applies to AI security products with more force, because the output is fluent and therefore harder to doubt. A model that summarises alerts convincingly while missing the one that mattered produces confidence without resilience.
The test stays the same: what bad outcome does this prevent, detect, contain, recover from, or prove? Judge an AI capability by the same evidence as any other control.
What would prove us wrong?
Claim: unknown
The program assumes machine-carried continuous defensive work improves measured outcomes.
- Evidence that AI-assisted triage lowers analyst vigilance enough to offset the coverage gain, a well-documented automation effect in other fields.
- Results showing that generated incident timelines contain errors at rates that damage legal and regulatory positions.
- Data showing that attackers gain more from the same capability than defenders do, shifting the correct investment toward recovery rather than detection.
Patterns and anti-patterns
- Machine prepares, human decides (pattern)
- Accepting a fluent AI report as an assessment (anti-pattern)
Evidence for this program
- MITRE ATLAS, adversarial threat landscape for AI systems Relevant as AI components enter both the defensive stack and the attack surface.
- NIST AI Risk Management Framework Used when assessing AI components in defensive workflows, where the same evidence standard applies.
Research notes
Let machines watch; keep consequential decisions with people
Machines can watch systems and prepare evidence. People remain responsible for decisions that cannot be reversed. This note recommends that machines take the high-volume, repetitive work where errors can be corrected, and that the line between machine and person move only when measured performance supports it.
Fluent output is not evidence
A generated security assessment reads better than a human one and is harder to doubt. Readability and correctness are unrelated properties. The risk is that AI output raises confidence faster than it raises capability, so every generated claim should carry its source record.
Adversarial simulation may replace part of annual training
Generated, varied, current scenarios delivered continuously and measured behaviourally are a plausible substitute for a yearly module. Plausible is not proven. Three things would need to be true, and all are testable: simulation measures predict real behaviour, simulation does not damage reporting culture, and scenarios stay within ethical and employment boundaries.