Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

Palo Alto finds 37% high-severity AI exposures
09/22/26, 05:47 PM

Palo Alto Networks says Unit 42 Continuous Frontier AI Defense found exposures in 100% of assessed customers, with 37% rated high or critical across more than 100 customer engagements. The launch announcement describes a worldwide annual subscription using a multi-model harness; these are vendor-reported results, not an independent safety score. For operators, Microsoft’s 90% security-work routing pattern points to the near-term move: widen triage coverage with models, but keep patch approval and local replay tests.
Wire

Coherent launches PhotonLink for AI data centers
09/22/26, 04:55 PM

Coherent launched PhotonLink, an integrated-optics platform for AI data centers, with more than 10 CPO and more than 10 NPO customer engagements. Its launch release says it also has more than five chip-to-chip engagements and expects PhotonLink revenue to ramp in Q4 2026. For cluster builders, optical connectivity is moving from component sourcing to qualification queue; our analysis of NVIDIA’s photonic switch explains why headline bandwidth still needs port-level tests.
Wire

Teradata turns Tera into an agentic coworker
09/22/26, 04:52 PM

Teradata says its new Tera Harness supported 512 concurrent agents on one 8-vCPU VM, serving 279 tool calls per minute, in a company test. Teradata’s announcement introduces Tera Context Engine, Tera Harness, and Agent Skills, but says all three arrive in Q4 2026. That makes it a watchlist candidate, not a switch-now verdict; our benchmark of coding-agent harnesses shows how execution layers can change outcomes even when the model stays constant.
Wire

Snorkel AI raises $350M at a $3.5B valuation
09/22/26, 04:49 PM

Snorkel AI has raised $350 million at a $3.5 billion valuation, CEO Alex Ratner told Reuters in a report carried by CNA. The company says its annualized revenue run-rate has also crossed $350 million, up from roughly $20 million a year earlier, as it shifted toward finished datasets and reinforcement-learning environments for frontier labs. For builders, our analysis of automated alignment research costs is the adjacent warning: specialized data and evaluation work is becoming a real line item, not free research overhead.
Wire

Multiverse puts 2.9M models in a wind tunnel
09/22/26, 02:03 PM

Multiverse Computing launched Luminary, an early-access evaluation platform for full multi-turn AI workflows, against a market of roughly 2.9 million public models, the company says in its launch announcement. Its digital twins score cost per completed task, success, latency, and policy compliance across simulated conversations rather than single prompts. Builders choosing an agent model should file it beside Multiverse’s earlier model-compression work and demand task-level evidence before paying for frontier capacity.
Wire

US proposes one AI incident channel with China
09/22/26, 02:00 PM

Washington proposed 1 bilateral AI-incident notification mechanism with Beijing after Treasury Secretary Scott Bessent and Vice Premier He Lifeng met in New York, with the trigger limited to incidents reaching a national-security level, Reuters reports in its account of the proposed AI-safety mechanism. China’s public readout acknowledged AI discussions but did not confirm acceptance, so this is a proposed crisis channel, not a working hotline. Builders shipping frontier or agentic systems should file it beside the archive’s federal-versus-state AI governance analysis and watch whether incident reporting becomes a deployment requirement.
Wire

Cisco Talos finds 20 more AI malware samples
09/22/26, 12:53 PM

Cisco Talos says its open-source CAIRN framework found about 20 additional examples of AI-integrated malware, including one Windows family whose autonomous command system consulted four large language models. Talos’s CAIRN release and its CLOSEDQUORUM report say researchers could not confirm real-world use, even though the sample was built to steal credentials and cryptocurrency; WIRED’s coverage frames the finding as an early defensive signal, not proof of a broad attack wave. Builders should inventory AI-facing endpoints and require human-reviewed incident evidence before treating model-connected malware as an operational trend, extending the archive’s analysis of a leaked cyber model.
Wire

Meta patches Muse's one zero-day
09/22/26, 12:50 PM

Meta patched a zero-day in its Muse macOS AI agent after security researcher Patrick Wardle showed that a local attacker could take control of the assistant; The Verge’s report says Meta issued a hotfix within hours, called practical risk low because local access was required, and saw its stock climb 11% on Monday. Ars Technica’s assessment says the design granted Muse unusually broad privileges, while no real-world exploitation was confirmed. Builders should treat desktop agents with account and device access as privileged software, requiring scoped permissions, signed updates, and a kill switch before autonomous actions—an extension of the archive’s local-agent hardware analysis.
Wire

TacnIQ.ai raises $3M for tactile AI
09/22/26, 11:56 AM

Physical-AI startup TacnIQ.ai secured $1.5 million from In Group Holdings inside a $3 million pre-seed round to build tactile foundation models, according to EET Asia’s financing report. Its deployments have accumulated more than 5,000 hours of tactile interactions across logistics, construction, e-commerce, hospitality, and healthcare, while paying customers supply field data. Builders evaluating Enigma’s interface-first robot bet should treat this as a data-collection signal, not proof of general-purpose dexterity: tactile models become deployable only when sensor coverage, labeling, and real-world feedback survive beyond lab demos.
Wire

Primo raises €6.9M for IT agents
09/22/26, 11:56 AM

Paris-based Primo raised €6.9 million ($8 million) in a round led by Headline and Global Founders Capital to put autonomous agents into IT operations, according to coverage of Primo’s financing. The company says it serves more than 400 customers and has passed $4 million in annual recurring revenue, with agents handling onboarding, access provisioning, and offboarding across devices and SaaS. Builders should file Primo alongside Temporal’s durable-execution cost gap: agent ROI depends less on chat quality than on whether system context, permissions, and recovery controls make autonomous actions auditable.