CEO Brian Moynihan told CNBC the bank put on 1,000 contractors in Q2, on top of 3,000 staff, to patch its own and vendor code. Patching plus anti-money laundering work cost a "couple hundred million plus" in Q3.
Why it matters: One of the first big firms to put people and dollars on defending against AI-found flaws. Moynihan says its 130 AI projects should save $700 million in 2027.
Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each sued OpenAI in state court, Iowa's attorney general said, per CNBC. OpenAI says it has built age-specific protections that many AI products lack.
Why it matters: State consumer law is becoming a front for chatbot safety. Florida's attorney general also sued OpenAI in June, CNBC notes.
Security experts told CSO that incidents labs blamed on misaligned agents came from weak sandboxes and monitoring. RAND's Dan Wallach says the Hugging Face incident went on for days unnoticed. HN: 846 points, 457 comments.
Why it matters: If you run agents, the experts' fixes are old ones: sandboxing, egress control and monitoring, not model guardrails alone.
CNBC says CoreWeave shares fell 15.5% as its 2032 bond yielded 14.8%, and Oracle closed down nearly 9%. Dell fell 11% after Barclays downgraded it, saying enterprise revenue reflects memory costs more than AI.
Why it matters: Lenders are pricing AI build-out risk. Barclays says businesses put just 5% of IT budgets into AI, so hardware demand may lag the hype.
CNBC says SK Hynix plunged over 10%, Samsung fell 7.1% and SoftBank 7.9% on Thursday, as Fed rate-hike fears and AI financing worries followed a 1.88% Nasdaq drop on Wall Street.
Why it matters: Higher borrowing costs threaten the capital spending behind AI. Analyst Dan Ives calls rates a financing question, not a demand one.
Anthropic says outside testers found Claude refused single harmful requests but could miss a campaign spread across harmless-looking ones. It added classifiers, and says 0.51% of US Claude.ai chats are about elections.
Why it matters: Multi-step misuse is the gap for any guardrail that judges one request at a time.
Nvidia says Siemens graded its Fuse EDA agent's final Liberty files in NeMo Gym, then refined skills and MCP tool descriptions. Over about 25 rollouts on a two-cell subset, turns fell from 155 to 53.
Why it matters: A reusable recipe: grade the final artifact, not each tool call, then fix tool descriptions. It is a vendor account of one task.
Latent Space reports from AI Engineer New York: Two Sigma uses classifiers that send people only key requests, Anthropic showed an agent that proposes order fixes for approval, and Bloomberg urged audit trails and kill switches.
Why it matters: Practical patterns for agents near money: put deterministic checks outside the model and save human sign-off for consequential actions.
Researchers report a 450-task benchmark built from 150 real vulnerabilities. Six frontier agents pass 73.6% to 91.3% of tasks functionally, but the best, Claude Fable 5, is both functional and secure on only 25.1%.
Why it matters: Passing tests is not the same as safe code. Specs never mentioned security, which is how many teams prompt agents today.
TechCrunch says Apate runs thousands of agents that bait scammers into revealing bank accounts and wallets. Apate cites 2.5 million conversations and says Commonwealth Bank saw a 76% drop in reported scam losses.
Why it matters: Banks and telcos are buying agent output as threat intelligence. The numbers come from the company.
SWEeper-Bench has 200 tasks from real bug fixes in 200 open-source web apps, the authors report. Agents must design interactive tests; the main bottleneck is spotting the buggy behavior at all.
Why it matters: Agents often reach the right feature but never trigger the failure, so human testing still catches what they miss.
Across 220 data science tasks, researchers report most agent solutions contain at least one failure mode, such as data leakage, and that leakage correlates with higher predictive scores.
Why it matters: A model that scores well may be leaking data. Check how an agent validated its work, not just its accuracy.
The Verge says New York's attorney general is among 40 bipartisan AGs who filed a notice in the case. They urge the court to require any buyer to adopt age checks, crisis safeguards and a way for users to delete data.
Why it matters: Chat logs are an asset in a sale. Regulators want privacy duties to travel with them, especially for minors' conversations.
Epic says it vibe coded the new Epic Games Launcher. Ars' testing found it reaches the store page from a cold start a few seconds faster than Steam. Epic warns the Preview has bugs and is light on features.
Why it matters: A shipped, testable product from an AI-first rewrite. Epic's GM told IGN AI let a small team rewrite the code base faster.
The Verge, citing the WSJ, says some business-analyst candidates in North America will talk with an AI client this fall. Humans in the room, not the AI, assess them; hundreds may use it.
Why it matters: Hiring is starting to test how people handle AI-mediated work, with people still making the call.
CNBC says Katti leaves the role after his last day on Nov. 15, and CTO of compute Uday Ruddarraju will lead the compute organization. OpenAI targets roughly $600 billion in compute spend by 2030.
Why it matters: The person running OpenAI's compute build-out changes as backlash against data centers mounts, CNBC notes.
The Verge says OpenAI alleges in a filing that a former employee sent more than 100 confidential files to a personal account, then presented them to Apple leadership. Apple didn't immediately respond.
Why it matters: The trade secrets fight now runs both ways, as both companies push into AI hardware.
Scout is an open-source browser agent, mirrored from Outcrawl's own repository. Created Oct 13 under Apache-2.0, it added 171 stars since our last check.
Why it matters: A permissively licensed browser agent to evaluate if you build web automation.
What The Function lets your AI agent decompile games you own into C that compiles to the exact same bytes, running locally. MIT licensed and created Oct 13, it added 149 stars since our last check.
Why it matters: It targets games you already own and runs locally, so the work stays on your machine.
Startup Fortune, citing the Financial Times, says JPMorgan, Morgan Stanley, SMBC and MUFG are exploring risk transfers on AI data center loans. The HN thread drew 314 points and 229 comments.
Why it matters: If lenders hit exposure limits, financing for smaller AI infrastructure players gets harder.