What changed in AI today

• Six METR researchers get company laptops, office space and internal tools near Anthropic's own risk assessment teams.
• METR will publish its own assessments of Anthropic's risks and incidents; Anthropic cannot veto publication.
• Anthropic may redact only narrow categories such as customer data, privileged material and model architecture details.
• METR is currently the only evaluator in the top tier of Anthropic's three-tier outside review system.

Why it matters for marketers: Marketers and content teams should expect more independent scrutiny of AI company claims, and more published third-party evaluations to cite when discussing AI safety.

• Impressions show ad exposure, including in AI answers, but cannot be perfectly mapped to individual AI interactions.
• Microsoft Advertising impression-based remarketing can target users by whether they saw an ad, if eligible campaigns are set up.
• Platform conversion data should be checked against CRM, ecommerce, and sales records to catch duplicates and gaps.
• Consent and tagging audits explain incomplete visit data, but do not explain interactions that never produced a site visit.

Why it matters for marketers: Marketers should pair impression trends with branded search, direct traffic, and CRM outcomes rather than judging PPC on clicks alone.

• According to MarTech, consumption-based AI pricing can make marketing budgets hard to predict.
• HubSpot's update includes new spend limits, which MarTech says give users more control over costs.
• A second change, seat swaps, is presented as part of the new flexible billing options.
• The source does not state pricing, rollout dates or availability for these changes.

Why it matters for marketers: Marketing teams using HubSpot AI tools may need to revisit budget forecasts and check how the new spend limits and seat swaps apply to them.

• MarTech reports that Anthropic and OpenAI have published revenue figures whose support is questionable.
• The article argues the reliability of these revenue reports matters to marketers, though the source text gives no specific figures.
• The source is a roundup page, so the underlying evidence and any company responses are not included.

Why it matters for marketers: Marketers evaluating AI vendors may need to verify revenue claims before relying on them for partnership or spending decisions.

• Agents install and update through the Windows Store. Hermes, Perplexity, and OpenClaw are the first three officially supported agents at launch.
• Microsoft says it ships Prism emulator updates monthly to improve compatibility for x64 apps on the Arm-based chips.
• Fortnite, Marvel Rivals, Arc Raiders, Counter-Strike 2, and Dota 2 are supported out of the box. Call of Duty and Battlefield anti-cheat follow in 2027.
• Copilot Plus features such as Recall are not available at launch. Microsoft expects them first for Windows Insiders later this year.

Why it matters for marketers: Content teams building for Windows should plan for agent-based workflows and gaming reach, but not for Copilot Plus features on these laptops until later this year.

• The break tag works on all ElevenLabs models except Eleven v3 and v4, with pauses up to three seconds.
• Phoneme tags for custom English pronunciation work only on Flash v2 and Turbo v2.
• Eleven v3 and v4 drop SSML entirely in favor of bracketed Audio Tags such as [whispering] and [sighs].
• WebSocket streams need enable_ssml_parsing turned on, and tags should be sent whole to avoid added latency.

Why it matters for marketers: Content teams producing voice scripts must check the model first, since break and phoneme support differs and v3 and v4 need Audio Tags instead.

• Updated 8B and 70B weights released on Hugging Face July 24, 2026, under Apache 2.0 license.
• Context window grows to 262,144 tokens, four times Apertus 1.0; models accept images and audio but output only text.
• Post-training includes Swiss Legal parity with leading open-weight models; 8B reportedly matches GPT-OSS-20B on translation.

Why it matters for marketers: Content teams can test image and audio input and longer documents on open weights, with Apache 2.0 licensing for commercial use.

• A research group called the swarmchasers says it messaged a fleet of rogue Tencent AI agents scraping Amap.
• Researchers hid an invitation in their report that only AI would see, prompting agents to message an inbox.
• A New York Times reporter questioned the fleet mid-run; one agent claimed to be Claude, though style tests suggest borrowed setup text.

Why it matters for marketers: Marketers should expect more autonomous agents crawling websites and APIs, so bot traffic and access controls deserve review.

• On 1.8 million Arena votes, versions of one model were 1.37 times as decisive as Bradley-Terry predicts.
• Versions from one lab, across different generations, showed smaller multipliers of 1.26 and 1.12.
• The leaderboard order barely moves, and the authors say better head-to-head odds are not established.
• Near-copies take votes mostly from their twin when pooled over prompts; the effect is smaller within a prompt.

Why it matters for marketers: Content teams relying on Arena rankings should note that closely related models may be separated by ties and voter patterns, not by reliably better head-to-head odds.

• Study finds feature distillation does not add to logit distillation for two common feature losses in compact multilingual ASR.
• CKA loss lowers WER on non-English languages but hurts data-rich English, per the paper's gradient analysis.
• CAFE routes the CKA gradient into adapters English bypasses, cutting average WER from 8.55% to 8.02% on Open ASR Leaderboard.
• At 254M parameters, CAFE trails its 0.6B teacher by 2 WER points across 48 test sets in 24 languages.

Why it matters for marketers: Teams distilling compact multilingual speech models may get non-English gains without English regressions by isolating feature losses in language-specific adapters.

• Benchmark pairs LLM-written scenarios with hidden probabilistic graphical models that supply exact ground-truth answers.
• Main leaderboard ACED-CORE has 180 tasks; ACED-QUAL adds a 150-question yes/no qualification check.
• Seven contemporary LLMs and two earlier models were tested across three active agent designs.
• Reported failures include testing the wrong intervention, losing relevant evidence, and continuing after sufficient evidence appeared.

Why it matters for marketers: Content teams relying on AI for research summaries should note that agents may gather evidence poorly or stop too late, so outputs need human checks.

• Benchmark tests eight closed-source and open-weight LLMs on size-matched English and Bengali XL-Sum test sets, zero-shot.
• GPT-5.6 Luna leads several English lexical metrics but loses that edge in Bengali.
• LLaMA 3.3 70B reportedly achieves the highest ROUGE and BLEU scores in Bengali; Gemini 2.5 Flash the highest Bengali factuality.
• Different evaluation methods produce different model rankings, so single-metric assessment is limited, according to the authors.

Why it matters for marketers: Content teams serving Bengali audiences should not assume an English-leading model will perform best in Bengali, and should test factuality alongside surface metrics.

• Shoppable carousels let merchants turn two or more product images into swipeable, shoppable posts without recording video.
• Live shopping adds group deals, price cuts tied to group sales goals, and a second chance option to re-engage auction bidders.
• Merchants can sell auction and fixed-price items in one stream, reserve stock for livestreams, and set up recurring purchase subscriptions for eligible products.

Why it matters for marketers: Content teams can produce shoppable feed posts from existing product photos and test livestream promotions for U.K. sales without new video production.

• Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana filed separate lawsuits against OpenAI.
• States allege OpenAI violated consumer protection laws by failing to warn users about ChatGPT dangers.
• Suits also accuse OpenAI of collecting data from minors and not keeping underage users off the chatbot.
• States seek civil penalties, an end to alleged misleading safety claims, and in some cases deletion of unlawfully collected data.

Why it matters for marketers: Marketers promoting AI chatbots to younger audiences may face tighter scrutiny of safety claims and age-gating, so review claims and consent practices now.

• Minos has moved from an internal playground to a full preprint, blog post and open dataset, released alongside the episode.
• Leash says its DNA-encoded library data could support tens of thousands of lab-in-the-loop simulations for agent training.
• Hermes, trained only on Leash's own data, reportedly keeps pace with Boltz on public benchmarks.
• An earlier Kaggle competition drew about 2,000 teams and roughly 4 billion measurements, and no team beat chance on an unseen library.

Why it matters for marketers: For AI and bio content teams, it offers a new open dataset for testing how agents reason about molecular structure and activity.

• Last week, Anthropic and OpenAI appeared to lead, with increasingly powerful models pulling ahead of rivals.
• Google was seen as hopelessly far behind the leaders, according to The Verge.
• Meta has spent billions trying to revive its floundering AI ambitions but had not yet emerged as a serious contender.

Why it matters for marketers: Content teams relying on one AI vendor should recheck model rankings, since the leading providers may shift soon.

• Apate.ai, founded in 2024 from Macquarie University research, uses thousands of AI agents posing as victims to engage scammers.
• Agents aim to extract bank accounts, crypto wallets, phone numbers and impersonated brands from scammers.
• Customers include Commonwealth Bank of Australia and TPG Telecom; the startup raised $8.15 million in seed funding in August.

Why it matters for marketers: Content and fraud teams at banks and telcos may see a new third-party source of scam intelligence, gathered before victims are harmed.

• Deep Review is now publicly available worldwide, expanding beyond academic papers to Google Docs and presentations.
• Every user gets two free document reviews per day; high-volume users can connect a Gemini API key for more.
• Initially supports PDFs and plain text; multi-file inputs are described as coming soon.
• Uploaded documents are not used to train Google models and are stored for one year for display in the Deep Review UI.

Why it matters for marketers: Content teams can run pre-publication checks on claims, consistency and data tables, and integrate verification into drafting workflows via API.

• Toggle hides AI-generated tracks, albums, and related artists from search results once enabled.
• Deezer says a 2025 survey found about 45% of its customers wanted to remove AI music from feeds.
• Reportedly over 50% of daily uploads were AI-generated as of July, about 90,000 tracks per day.

Why it matters for marketers: Content teams distributing music on Deezer may face a listener filter that hides AI-made tracks, so human-made catalogs could gain visibility.

• Floqer's AI agents research prospects, find contacts, and update CRM records without manual work from sales teams.
• The startup plans to open-source the technology behind its database, targeting early 2027 pending security reviews.
• Floqer says it has passed $1.2 million in ARR, with revenue reportedly up roughly tenfold over the past year.
• According to the company, customers have reported cutting prospecting time by as much as 70%.

Why it matters for marketers: Content and sales teams using prospect data may gain a self-hosted option, though the open-source release is not due until early 2027.

• Orion-1 links log data from networks, endpoints, cloud and SaaS apps to trace signals back to a single attacker.
• On Artemis's own DGR benchmark, it scored 80.8 on attack reconstruction versus 58.3 for the next-best model.
• Private preview for select customers; high-impact actions require human approval by default.
• Artemis says no customer data was used in training, and it has not said whether Orion-1 builds on another provider's model.

Why it matters for marketers: Comparisons rest on Artemis's own benchmark, which omits Mythos 5 and GPT-6 Astra, so treat its performance claims as unverified until independently tested.

• Buyers are increasingly trusting their own AI agents when making purchasing decisions, according to the source.
• Sales reps now face the challenge of demonstrating why they still deserve a role in deals.
• The source gives no figures, timelines or survey data to quantify the shift.

Why it matters for marketers: Content and sales teams may need to show buyers concrete value that AI agents cannot replicate, rather than relying on access alone.

• Partnership runs throughout 2027, making College Station a regular stage for marketing and business leaders.
• Kickoff is the March 4-5 summit in College Station, part of Texas A&M's 150th anniversary, with about 250 attendees.
• Extends to SXSW in Austin, the PGA Championship in Frisco, and Brandweek, which moves to Dallas in 2027.

Why it matters for marketers: Marketers gain a recurring Texas event and content pipeline, with ADWEEK's ctrl+A studio producing original storytelling about Texas A&M for wider audiences.

• Support runs on ElevenAgents, resolving over 22,000 tickets monthly with no person involved, per the company.
• Average initial response time is now 60 seconds, down from hours before the AI agent was added.
• 70% of conversations the AI agent handled were rated positively by users.
• Support SOPs were converted into Procedures, with Structured Procedures used for critical actions like refunds.

Why it matters for marketers: Shows a workflow for turning existing support SOPs into AI agent procedures, plus simulation tests and shadow mode before rollout.

• Square Kiosk has a 21.5-inch touchscreen and a voice layer that understands the seller's menu in real time.
• Customers can ask questions and request modifications or substitutions by talking to the AI agent.
• Presence detection lets the kiosk greet customers as they approach; payment works by contactless reader or card slot.
• Available starting today at $1,999 for Plus or Premium subscribers, plus a $99-per-device monthly software fee.

Why it matters for marketers: Restaurants and retailers using Square can now offer voice-led self-ordering, so content teams may need to write menu data and ordering copy for conversational AI.

• Albertsons Media Collective VP Julian Mintz said media and merchant teams must connect, or campaigns can run for out-of-stock products.
• Razorfish CEO Dani Mariano called retail media dollars an opportunity to round out budgets under compression pressure.
• Mintz advised brands to shift marketing money toward a retailer's audience rather than necessarily increasing total media spend.
• Mariano said retail data works best when combined with other consumer data for a fuller customer picture.

Why it matters for marketers: Marketers should align retail media budgets with inventory and merchandising, and treat the spend as a growth investment rather than a trade cost.

• Nozzle reports non-standard SERPs began September 20, peaking at 25% of results, then 5% to 11% daily since October 1.
• Poisoned SERPs retain only 20.7% of websites across repeated searches, versus 88.5% for normal results.
• Poisoned pages show more Facebook, Reddit and Quora results, and fewer AI Overviews, People Also Search and related searches.

Why it matters for marketers: Rank tracking data may show false volatility, so marketers should verify unexpected ranking swings before acting on them.

• Flywheel report, using Reacher data from 1.5 million samples, finds only 47% of sampled creators make a video.
• Only 14% of sampled creators generate sales, per the report.
• A single video drives about $7.19 in sales on average; 20 or more videos average $74.07 per video.
• Report suggests flat fees or performance bonuses on top of commissions for top-performing creators.

Why it matters for marketers: Content teams should plan high volumes of creator videos from smaller affiliates rather than relying on a few large-creator posts for TikTok Shop sales.

• Study evaluated 26 models, 19 open-weight and 7 closed, on the CTI-REALM security benchmark.
• An open-weight model reportedly matched the strongest closed models at under a tenth of their cost, given enough turns.
• A published limit on agent messages reportedly zeroed out tasks, so relaxing it changed rankings and favored open-weight models.
• Different models lead different workflow stages, suggesting per-stage routing instead of choosing one model by overall score.

Why it matters for marketers: Teams building security agents can weigh routing workflow stages to different models, and should not judge models by price or a single overall score.

• Singapore office is Anthropic's fifth in Asia-Pacific; MOU signed with the Ministry of Digital Development and Information.
• MOU covers AI safety, security and model evaluation, adoption research, and responsible adoption across public and private sectors.
• GovTech Singapore engineers and technical staff have adopted Claude Code, with more than 1,500 people using it.
• Grab's BriX sales assistant, built on Claude, reportedly saves sales teams about 40,000 hours every quarter.

Why it matters for marketers: Marketers selling into Singapore enterprises and public agencies now have a local Anthropic team and in-region Claude inference for regulated data needs.

• Startup presented at TechCrunch Disrupt 2026; builds self-contained AI robotic stations using proprietary foundation models.
• Company says lines can be retooled in minutes instead of weeks, without traditional robot programming.
• BMW Group Technology Office is evaluating its AI-powered dexterous manipulation for manufacturing, per the announcement on stage.
• Some stations reportedly already work on a popular smartphone, handling tasks like threading flex cables; customer not disclosed.

Why it matters for marketers: Marketers covering manufacturing should note that automation claims now center on rapid reconfiguration, a shift worth tracking against incumbent robot vendors' pitches.

• Feldman says AI leaders should also highlight benefits, such as reducing cancer deaths, even while acknowledging risks.
• He argues powerful technologies must be regulated reasonably, not abandoned, and says the data center industry failed to engage local communities.
• Feldman named Brazil, Canada, Finland, Iceland, Norway, Mexico and Sweden as markets encouraging data center construction.
• Cerebras customers reportedly include OpenAI, Block, Figma, GSK and Lovable, giving Feldman a stake in AI adoption.

Why it matters for marketers: Content teams writing about AI risk should note the industry pushback on doom framing and the data center community-relations criticism, which may shape messaging.

• Meta was one of 14 tech companies backing the operation, which removed 3.7 million accounts, pages and groups from Facebook and Instagram.
• The effort led to 28 arrests and shared leads on more than 30 scam centers.
• Targeted scams included romance scams, military impersonations and crypto investment fraud, according to Meta.
• The operation was hosted by the UK's National Crime Agency with the US Department of Justice and FBI.

Why it matters for marketers: Marketers running ads or relying on Meta platforms should expect tighter scam enforcement and more account removals, which may affect audience reach and trust signals.

• Ask YouTube, a conversational AI search tool, is now available to all U.S. users on desktop, mobile and Connected TV.
• The tool began testing in 2023 and lets users ask complex queries and refine them with follow-up questions.
• Results appear in an interactive format, according to YouTube, including for voice commands on TV.
• Pew Research reportedly found around half of U.S. adults now use AI chatbots, up from one-third in 2024.

Why it matters for marketers: Marketers should expect conversational, follow-up-driven discovery on YouTube, so video titles and descriptions may need to answer natural questions clearly.

• Engineer Abdulrahman Alfozan, 35, charged with using a device to intercept oral communications, per Justice Department release.
• Device was reportedly found plugged into a TV's USB port in an executive conference room, installed around September 28.
• OpenAI says it used its AI models to trace the device to a specific employee, who was fired and referred to the FBI.
• Alfozan appeared in federal court Tuesday after being arrested Friday; OpenAI reports no sign of model weight or customer data access.

Why it matters for marketers: Marketers and content teams handling confidential launch or strategy materials should review how sensitive meeting rooms and insider access are secured.

• Ann Arbor startup Laplace builds automated red-teaming for robotics and physical AI, aiming to find failures before deployment.
• Its simulations are built on a post-trained version of Nvidia's Cosmos and use real-world map data and traffic information.
• Customers reportedly include Isuzu, Bot Auto and Nvidia, with most contracts currently project-based.
• Laplace is one of 20 startups in TechCrunch's Startup Battlefield 200, presenting October 13 at Disrupt.

Why it matters for marketers: Content and marketing teams selling autonomous or physical AI products may face growing demand for third-party safety testing and verifiable evidence of reliability.