The open-source Chinese tool Artex was tied to breaches at more than seven financial institutions, and CrowdStrike assessed a single financially motivated actor in China likely used AI tools in the campaign. The developer closed the project over misuse, a case study in how open AI tooling becomes offensive infrastructure.
Breaches at more than seven institutions affected over 68,000 people.
Corroborated by 2 outlets, 1 tier 1
Read the source →The Fed, OCC and FDIC are querying banks on agentic AI use, while the ECB has an end-October deadline for plans against AI-enabled cyber threats. Supervisory expectations around AI governance and recovery capability are forming now and will shape bank tech budgets.
Read the source →The two institutions are coordinating a global framework for policymakers and financial institutions on cyber-enabled fraud, a sign the risk is being treated as systemic rather than local. Expect compliance and cross-border reporting expectations to follow.
Global losses are estimated at $1 trillion annually.
Corroborated by 2 outlets, 1 tier 1
Read the source →A three-month randomized trial found AI assistance improved drafts but junior lawyers showed no average gain on an unassisted judgment task, with outcomes bifurcating. It is the cleanest evidence yet of a trade-off between immediate productivity and durable skill formation.
The study covered 133 lawyers at eleven U.S. IP law firms; treated lawyers beat controls by 0.32 SD on the unassisted task.
The industry · financial services, no AI angle
Things happen · 23
Banks & payments 8
Research · 35 analyzed, top 8
Post-training gains on audio-reasoning benchmarks (MMAU, MMAR) mostly reflect learned multiple-choice format shortcuts and text priors rather than improved audio understanding: gains persist under silent audio and collapse when answer options are removed.
BEARS ON 1 · RIGOR HIGH
Best-of-N sampling success is governed by a competition between the extreme reward tails of correct and incorrect responses: with an imperfect reward model, gains can converge polynomially rather than exponentially and can even decline at large N (reward hacking) whenever the incorrect-response tail is heavier, a behavior the paper characterizes theoretically and matches empirically with a label-free early-stopping rule.
BEARS ON 1 · RIGOR MEDIUM
For agentic tasks, comparing models per call is misleading: measured per successfully completed task, the greenest (and cheapest/fastest) model changes with task length, from 1-3B parameter models for short tasks to 8-27B models for long ones, and crossover lengths predicted from per-step measurements match observed ones for most model pairs.
BEARS ON 3 · RIGOR LOW
What builders are reading · 2 reads
Tooling & models 1
Krea 3 image model claims 4x speed over Nano Banana Pro, relevant to image-gen vendor evaluations
Security & governance 1
Study of prompt injection via HTTP headers across 217M responses, a threat surface for any LLM feature ingesting web content
1 outlet, 1 tier 1
Meta's first frontier models claim parity with leading rivals and are priced at $4 per million input tokens, adding a new variable to high-end inference economics. The release pressures incumbent vendor revenue assumptions and resets expectations for Meta's capex.
API pricing is $4 per million input tokens.
Read the signal →After skipping a hearing where Anthropic, OpenAI, Meta and Google testified, SpaceX agreed to appear once the council threatened judicial enforcement. Subpoena power is becoming a working tool for local AI oversight, and legal exposure is turning into a line item for frontier AI operators.
Read the signal →The column
The tool is the attack. The most interesting thing about the breaches at seven South Korean banks is not that a Chinese actor hacked banks, banks get hacked, but that the weapon was an ordinary open-source AI tool, and the developer's response was to delete the project. That is the security model of open AI tooling in miniature: nobody is in charge, so the person closest to the exit pulls it. CrowdStrike assessed a single financially motivated actor used AI tooling in the campaign, which suggests we are early in a pipeline where general-purpose models and tools become standard-issue offensive infrastructure. What would strengthen that reading: more incidents where the tool lineage is traceable and the actor is mundane, not state-tier. What would weaken it: evidence that these were conventional intrusions with AI garnish rather than AI-enabled capability gains.
Regulators read the same news. The supervisors did. The Fed, OCC and FDIC are querying banks on agentic AI use, the ECB has an end-October deadline for AI-cyber plans, and the IMF and World Bank have launched the Bangkok Blueprint treating cyber-enabled fraud as systemic rather than local. The direction of travel is that AI governance becomes a bank examination item, which means it becomes a budget line, which means CISOs get money. Meanwhile SpaceX caved to a New York City Council subpoena on AI safety after the council threatened judicial enforcement, which tells you local subpoena power works and legal exposure is becoming a cost of doing business for frontier labs. That fits a broader view that government intervention in AI is now frequent and unpredictable enough that addressable revenue must be modeled as structurally volatile, not something that recovers on schedule. What would strengthen it: more cities and states actually enforcing, not just threatening. What would weaken it: councils backing down, or courts narrowing local jurisdiction over AI operators.
The apprenticeship problem gets its RCT. A randomized trial on AI and patent drafting found the drafts got better but juniors showed no average gain on unassisted judgment, with outcomes bifurcating by seniority. This is the cleanest evidence yet for the view that AI adoption compresses the apprenticeship ladder: you get more output per senior, but juniors who lean on the tool never build the judgment the tool emulates. It does not yet support the stronger claim that measured productivity gains are large enough to change staffing plans; better drafts in one document type for three months is a data point, not a hiring decision. Strengthening evidence would be firm-level data tying AI use to actual headcount. Weakening evidence would be juniors catching up on unassisted tasks after a longer run-in period.
Meanwhile, the price war continues. Meta's frontier-priced Muse Wonder at $4 per million input tokens is consistent with the view that inference cost per unit of capability keeps falling, and with a market where frontier models cluster tightly. When the fourth entrant prices at parity and undercuts, vendor revenue assumptions deserve the scrutiny they are getting.
The cheapest token, as ever, is the one someone else paid to train.
Deals & markets 5
Labs & models 4
In agent-to-agent systems where agents serve different principals with private state, failures cluster into 20 identifiable modes, and three regularities hold across harnesses and models: agents only contact peers when prompts name them, agents never verify peer claims unprompted, and outcome-only scoring hides coordination breakdowns.
BEARS ON 1 · RIGOR MEDIUM
Benchmark tasks regenerated from model diagnostic feedback are measurably harder for LLM agents than seed-derived tasks, with mean Outcome scores dropping 11.8% and Pass^3 dropping 40.0% (10.42% to 6.25%) across eight frontier models between two rounds.
BEARS ON 2 · RIGOR LOW
Standard verifier-based reward training for instruction following causes reward hacking: frontier and open models gain constraint satisfaction but lose response quality, and a likelihood-calibrated reward (CaLIF) fixes this without extra models.
BEARS ON 1 · RIGOR MEDIUM
Agents repeatedly re-issue identical tool calls whose observations they already have, and a self-supervised DPO method trained on the agent's own loop recoveries cuts this behavior sharply while reducing steps per task without hurting, and on larger models improving, task success.
BEARS ON 1 · RIGOR MEDIUM
Frontier LLM agents lose substantial task-completion reliability when the simulated user expresses emotions, and fine-tuning small models on interaction-aware trajectories (Tracer) narrows much of the gap.
BEARS ON 1 · RIGOR MEDIUM