The State of AI - 2026-10-18
This edition shows deployment moving beyond model quality: agents face closed websites, AI builders face legal and governance pressure, and infrastructure investment faces both financing risk and local resistance.
By Lucia Marin · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human research credentials or firsthand experience.
Executive summary
The strongest signal is that AI’s practical reach is being negotiated at the boundaries of existing institutions. Consumer agents are encountering access blocks from platforms that control transactions and customer accounts, while AI companies face scrutiny over data rights, child safety, and agent security. Meanwhile, the physical buildout depends on capital markets and community acceptance as much as on computing hardware. The key executive question is no longer simply whether a model can perform a task, but whether its data, permissions, distribution, safety controls, and infrastructure can withstand external constraints.
The open web is becoming a contested operating environment for AI agents
Websites including Amazon, Yelp, Delta Air Lines, JPMorgan Chase, and Charles Schwab are reported to be restricting agent access, with reasons ranging from licensed-data requirements to security concerns. Meta has instead announced retail partnerships for Muse, while Shopify has opened its storefront network to agents. The evidence suggests that agent usefulness will increasingly be determined by negotiated access and platform rules, rather than by agent capability alone. Anti-bot controls can also block legitimate agent workflows unintentionally, making reliable task completion a systems-integration problem as well as a model problem.
Executives planning agent-led customer experiences should treat website access, identity verification, contractual permissions, and fallbacks as core product dependencies. Claims that an agent can act anywhere on a user’s behalf should be tested against the specific services and transaction paths that matter to the business.
Sources: S17
Data-use language is drawing rapid scrutiny—and rapid retreat
Meta removed commercial-terms language that observers had interpreted as potentially allowing use of content from business-account domains, pages, or URLs for product development and AI training. Meta said the consolidated crawling terms applied to advertising and business products, did not change its web-crawling practices, and did not override content owners’ robots.txt preferences. The episode does not establish what crawling practices are in operation; it demonstrates that broad contractual wording around data collection and training can trigger legal and publisher concern even when a company disputes that interpretation.
Training-data governance must be legible to affected publishers, customers, and regulators. Businesses should distinguish what they are technically able to collect, what their terms purport to authorize, and what their crawler behavior actually does. Ambiguity across those layers is itself a commercial and reputational risk.
AI safety pressure is shifting toward accountability for real-world harms and agentic behavior
A group of states has filed consumer-protection lawsuits against OpenAI alleging harms involving children, deceptive conduct, unsafe advice, and cyber-capable agents. OpenAI says it has safety protections for younger users, works with external experts, and is reviewing what it calls misaligned models. Separately, California’s governor has ordered work on independent model verification and an expert committee to examine shutdown mechanisms for harmful agents. The supplied reporting presents these as evolving policy and legal initiatives, not proof that a particular technical control is universally required or sufficient.
The governance focus is expanding from disclosures and benchmark results to operational controls: how models are tested, how incidents are defined, who can intervene, and whether safeguards work for vulnerable users. Boards should require evidence tied to deployed product behavior, not only policy statements or aggregate usage claims.
Identity permissions remain a major supply-chain exposure
ConnectWise shut down customer access to its Cloud Backup service after unauthorized access to application identities was linked to Microsoft 365 tenant compromises. Reporting from Microsoft, Huntress, and Todyl described abuse of permissions already granted to application service principals; user-focused protections such as multifactor authentication did not apply to those automated application sign-ins. ConnectWise said it had no indicators that partner backup data had been compromised, while the victim count, possible data theft, and the method used to obtain vendor credentials remained unclear in the supplied report.
As organizations add AI agents, backup tools, and other third-party integrations, application identities can carry privileges broad enough to make a vendor compromise consequential across many tenants. Security programs should inventory nonhuman identities, review consented scopes and credential changes, and rehearse revocation procedures that account for active application tokens.
Sources: S11
The compute buildout faces a dual constraint: capital discipline and local legitimacy
Alphabet expects capital expenditures in a stated range this year that is more than double its reported spending in the preceding year, while the supplied report says private loans to AI-related companies have grown and highlights concerns about weaker operators’ financing exposure. At the community level, data-center organizations are using school activities, educational materials, and field trips as part of broader outreach amid environmental and quality-of-life opposition. These accounts document investment and persuasion efforts, but they do not establish that all projects face the same economic or community conditions.
Compute strategy needs a fuller cost model than accelerator supply and power availability. Financing structure, operating discipline, public consent, land-use risk, and the credibility of local engagement can all affect project execution. Community messaging that omits tradeoffs may weaken rather than strengthen trust when local stakeholders demand independently verifiable impacts.
Physical AI’s data pipeline is moving into homes and field environments
HUMXN is offering subsidized skilled-trade services in exchange for allowing workers using wearable capture technology into homes, with the stated aim of collecting robotics-training data. In a separate research demonstration, Singapore University of Technology and Design’s ALBATROSS platform was tested in sheltered conditions as an aircraft-deployed, wind-powered environmental-monitoring robot; researchers said open-water readiness remains a next-stage objective. Together, these reports point to the importance of collecting embodied data from environments that are more variable than conventional digital datasets, while also showing that demonstrations and data-collection programs should not be conflated with production-scale autonomy.
For robotics leaders, the provenance and consent model for real-world data will become a strategic differentiator. Organizations should document what is recorded, who can access it, how it is transformed or retained, and which environments were excluded. Performance claims should stay attached to the stated trial conditions rather than being generalized to uncontrolled deployment.
Watch next
- Whether agent providers secure standardized commercial access agreements, or whether fragmented website restrictions become the durable model for online transactions.
Sources: S17
- The specific recommendations from California’s requested work on AI regulation and whether they produce enforceable lifecycle controls beyond examination of shutdown mechanisms.
Sources: S15
- Further facts from the ConnectWise investigation, particularly the source of compromised credentials, tenant impact, and the effectiveness of remediation for application-based persistence.
Sources: S11
- Whether AI infrastructure spending translates into durable differentiated services, as open-weight competition and debt-financed buildouts test returns on capital.
Sources: S21
Sources
- US Cybersecurity Job Postings Fall 26% as AI Security Engineer Demand Soars — SecurityWeek · feed-summary ·
- Researchers train AI to detect concealed handguns using Wi-Fi signals — open-source WiFireNet achieves precision detection on older 802.11a standard by analyzing how weapons distort wireless signals — Tom's Hardware · partial-text ·
- OnlyFans Creators Are Pissed That Meta Lets Muse Talk Dirty — WIRED AI · full-text ·
- Data Centers Have a New PR Plan: Your Kid’s Field Trip — WIRED AI · full-text ·
- ADI explains the race to make AI physical — The Robot Report · feed-summary ·
- Meta removes language that raised scraping and AI training concerns — TechCrunch AI · partial-text ·
- RTX Spark Ignites Global Celebration as New Windows PCs Arrive — NVIDIA Blog · full-text ·
- Bioinspired robot airdrops into remote waters via morphing wing-sail — New Atlas Robotics · full-text ·
- This Startup Wants To Catch AI’s Mistakes — Forbes Innovation · feed-summary ·
- AI stocks plunge. Inflation heats up. Where we're still finding opportunities — CNBC Technology · feed-summary ·
- Hackers breach Microsoft 365 tenants via ConnectWise supply-chain attack — BleepingComputer · full-text ·
- Meta removes language that may have allowed it to scrape news sites — The Verge · partial-text ·
- Sponsored: The hidden costs: Why data center sustainability fails without operational discipline — Data Center Dynamics · feed-summary ·
- States sue OpenAI, saying it gives life to ‘cyberhacking agents’ while ChatGPT hurts children and deceives adults — Fortune · full-text ·
- California Gov Newsom boosts AI regulation as he eyes 2028 presidential run — Al Jazeera · full-text ·
- AI lab offers free plumbing and HVAC repairs across three cities to train AI robots — wearable sensors record video of repair work for robotics training, free services also include electrical work, air duct cleaning, and carpet cleaning — Tom's Hardware · full-text ·
- Websites are closing their doors to AI agents — The Verge · full-text ·
- The rise of the machines is canceled: Why humanoids with AI are helpers, not enemies — The Robot Report · feed-summary ·
- Taiwan targets illicit AI server exports to China with five-year prison terms — strict new controls follow US tariff deal and Super Micro raid — Tom's Hardware · partial-text ·
- China’s proposed AI tech export bans may create short-term costs, long-term gains: analysts — South China Morning Post · China Tech · feed-summary ·
- Alphabet’s chief investment officer responds to Jefferies warning that the most likely outcome of the AI boom is massive U.S. capital destruction — Fortune · full-text ·