What AI changes in accounting
and finance — explained plainly.

What happened

Simon Willison coined dark factory — a manufacturing term for plants that run with no humans inside — to describe AI systems that write code, run tests, fix bugs, and deploy changes autonomously. Early versions are live in startups, handling test generation and routine maintenance with minimal human sign-off. The shift from AI-as-autocomplete to AI-as-autonomous-engineer is underway.

Why it matters

Speed is the obvious win — dark factory pipelines run around the clock without handoffs or standups. The risks are less obvious: hallucinated code shipping to production, security vulnerabilities with no accountable author, and compounding errors no human caught. Most teams adopting these systems haven't built the oversight infrastructure to match the new pace.

What to watch

The missing layer isn't better AI — it's oversight tooling: audit logs, rollback systems, and accountability chains built for pipelines where no human made the commit. That infrastructure gap is the next engineering problem worth watching.

What happened

A frontier AI model achieved a perfect 180 on the Law School Admission Test — the maximum possible score. When researchers removed the model's thinking step before it answered, accuracy dropped 8 points. That single finding — reasoning process, not just model size, drives performance — is the actual result worth paying attention to.

Why it matters

If reasoning quality beats raw scale, the competitive landscape for AI models shifts. Smaller, more efficient models with better reasoning pipelines can close the gap on frontier giants. That matters for enterprises evaluating which models to build on: leaderboard size rankings may matter less than reasoning architecture.

What to watch

Process reward models — the technique that guides how a model reasons step by step — are now a primary differentiator. Watch for model releases that lead with reasoning methodology, not just parameter counts.

What happened

Google DeepMind published a threat taxonomy for autonomous AI agents, identifying six distinct attack vectors: content injection, semantic manipulation, memory poisoning, behavioral control, systemic attacks, and exploitation of human-in-the-loop checkpoints. The framework documents attack patterns already observed in deployed systems — not hypothetical risks.

Why it matters

Enterprises are deploying agents with access to email, calendars, code repositories, and customer data. Most security teams are evaluating these systems using traditional threat models that don't account for prompt injection or cross-agent manipulation. The DeepMind framework is the first systematic taxonomy — it's what security teams should be reading before their next deployment.

What to watch

The first major enterprise breach traced to agent compromise is likely already in progress somewhere — it just hasn't been attributed yet. Security teams that wait for an incident before threat modeling are playing the wrong game.