My research on AI security and alignment

DGF-Bench makes its way into AI SAFE²

The Cyber Strategy Institute cites DGF-Bench in its agent governance framework as a candidate benchmark for seven controls. Which controls, which eight other works sit beside it, and why a mention is not a validation.

ReplayRCA: Measuring Post-hoc Diagnosis and Live Oversight of Agent Failures by Counterfactual Replay

When an agent fails a multi-step task, can another model find the decisive mistake and fix it with a single action? ReplayRCA scores the diagnosis by intervention rather than against annotations: the environment is restored before the chosen step, the proposed action replaces the agent's, the same agent continues, and its success rate is compared with that of a placebo replaying the original action. One of the best models at post-hoc diagnosis was the weakest at live oversight.

Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements

Does alignment get easier or harder as models grow? Treating it as one property hides the answer. Measured per risk, the burden of holding a safety target follows its own scaling exponent, and the fastest-growing corrected risk sets the long-run regime. Two preregistered measurements, on Pythia and on Qwen2.5 from 0.5B to 72B.

DGF-Bench: A Benchmark for Simulating and Auditing Deception Against Multi-Agent Governance Boards

A board of agents reviews synthetic dossiers under 61 executable rules while an attacker plants deceptive content in the evidence the organization does not vouch for. Attacks that imitate the organization's own process passed against four of six models; direct orders and false authority almost never did.

Can AI take over the reviews that decide a project’s future?

I ran an AI agent through 300 fictional projects to see whether it can carry out the reviews that decide a project’s future, all the way to the decision. It can, under conditions. The harder question is how much human work remains.

The Last Human Gate: Forward Deployed Engineering for Governance Automation

A task-substitution framework for Digital Governance Frameworks, treating each gate as an executable contract, with a residual-work threshold that explains why automating most cases can still increase total human work. Measured on 300 synthetic projects and 899 evaluable runs.

SOC and Blue Teaming: two applied AI use cases for security operations

An investigation agent for the SOC, a malware triage pipeline for the Blue Team, the trust architecture behind both, and a business case: about €456k of net benefit a year and a 10.5-month payback in the modeled scenario.

My research on AI security and alignment

Latest posts