How long your AI agent stays alive between resets is a product decision, and the data here says the long-lived ones quietly forget their instructions. This essay proposes a coordinator agent that lives a day, specialist agents that live about 30 seconds, and state written to ordinary files rather than held in the agent's head. It cites research where squeezing a conversation down to fit dropped standing rules in 30 to 59% of runs. Memory quality falls as it grows across every frontier model tested. Memory carried between sessions is also somewhere an attacker can plant text.
Claude is now the reasoning model behind Salesforce's Atlas Reasoning Engine. It also powers Agentforce Vibes and Agentforce Coworker by default, and is available in Agent Builder. Slack is now also powered by Claude. Salesforce for Claude is a plugin that gives sellers 37 prebuilt sales skills that work inside Anthropic's own interface rather than inside Salesforce's.
Cohere Parse is a cost-effective vision language model for processing large volumes of enterprise documents. It converts complex, multimodal files into structured machine-readable data. Parse detects and understands key visual elements and can process documents and images across nine major world languages. It can be accessed through the Cohere API for just $1.50 per 1,000 pages.
A fund that buys into the top slice of YC batches just published what it now reads into your Demo Day valuation cap. If you raised at one, an investor somewhere is running this math on you. The fund looked at 642 companies from the 2021 through 2024 batches. Companies priced above their batch median reached a Series A or beyond 25% of the time, against 8% for those priced below, and shut down at 6% versus 16%. The pattern holds inside each batch and among companies the fund passed on. Median caps went from $15M to $40M between W21 and P26.
AI-era businesses are far more distributed than ever before. They are increasingly forming in the outer suburbs rather than in city centers and high-density areas. The companies that are most responsible for AI development itself continue to cluster in major metro areas. AI-era business formation appears to be broadly distributed across the US, but major metro areas are seeing below-average business formation relative to their sizes.
No one wants a half-finished spaceship or just a shiny idea.
Listen Labs used a cryptic billboard in San Francisco to attract top engineering talent by challenging them to decode AI-related tokens.
Connecting with your customer and market and listening to your organic marketing efforts will enable you to see trends and opportunities before they show up in third-party data sets.
Bank Statement Converter is seeing more people canceling subscriptions and fewer people buying subscriptions, possibly because people are using AI to convert their bank statements.
Sponsored
Don't let rising GPU prices and thermal demands bottleneck your innovation. SiMa.ai's hardware/software co-optimized platform is purpose-built for Physical AI, delivering up to 10x better performance per watt than industry alternatives—without the GPU price premium.
Worried about switching costs? SiMa.ai's PALETTE NEAT [5] utilizes an agentic development workflow that analyzes your NVIDIA code repository and helps migrate your application logic. This automatically maps your codebase to SIMA.AI SILICON [6], preserving ~90% of your legacy software investment while eliminating CUDA dependencies.
Unlock best-in-class performance and escape vendor lock-in. Read the NVIDIA TO SIMA.AI MIGRATION GUIDE [7] today.
Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared driver spec that lets AI agents discover and operate lab and manufacturing instruments, microscopes, liquid handlers, robotic arms, and laser systems through a single interface instead of the bespoke, vendor-specific integrations that normally take weeks or months to build. The driver uses simple read/write primitives any device can implement, plus natural-language tags that capture physical characteristics (like a robot arm's weight) that used to live only in paper manuals or an engineer's head, giving the agent everything it needs.
Perplexity has partnered with Nvidia to launch Portable Computer, an agentic Computer platform designed to execute advanced AI agent workflows entirely on-device without continuous cloud API dependency or recurring token costs. Powered by Nvidia's compact edge silicon architecture, the system runs quantized local inference models to process multimodal search, complex reasoning, and persistent agentic tasks directly at the edge.
The NVM Express organization has released the NVMe 2.4 specification, integrating quantum-resistant encryption algorithms and advanced telemetry to address future-proofed enterprise security and datacenter energy constraints. These enhancements formalize a critical hardware-level bridge between next-generation flash controllers and high-efficiency datacenter storage architectures.
As high-bandwidth memory (HBM) architectures scale to 12-high and 16-high vertical DRAM stacks, semiconductor packaging faces severe physical limits across thermal dissipation, mechanical stress, and die warpage. Thinner DRAM silicon substrates required to maintain JEDEC package height profiles exacerbate through-silicon via (TSV) structural fatigue and increase susceptibility to micro-cracking during thermal cycling. The escalating complexity and defect density risks underscore that packaging and thermal reliability—rather than pure lithography—now dictate the economic scaling and memory bandwidth roadmap for frontier AI hardware accelerators.
ROS2SmolVLA is an open-source ROS 2 interface that deploys Hugging Face's lightweight SmolVLA model directly onto an industrial Universal Robots UR10e arm for local, vision-guided pick-and-place. If you have been playing with an SO-101 arm and would like to graduate up to controlling an industrial co-bots, this looks like a great starter solution.
NVIDIA just dropped a new tutorial on COMPASS, an agent-assisted residual reinforcement learning framework designed to adapt pretrained X-Mobility locomotion policies to different robotic platforms (such as Boston Dynamics' Spot). It offers developers a powerful methodology to reuse general navigation and locomotion skills across varied physical platforms, vastly reducing the time and cost required to train new hardware embodiments from scratch.
BigSky SF-2U870 is an enterprise-grade 2U rackmount development platform designed to accelerate RISC-V datacenter software validation and workload porting. Built around 32 high-performance P870-D cores clocked at 2.0 GHz, the system features full compliance with the RVA23 profile standard alongside support for 256 GB of DDR5-5600 memory, 64 PCIe Gen 5 lanes across four x16 slots, and dual U.2 NVMe storage. In addition to running enterprise distributions like Ubuntu 26.04 LTS and RHEL 10 out of the box, the platform can act as an accelerated host node, running Nvidia CUDA workloads directly on RISC-V hardware ahead of future NVLink Fusion integration.
Robotiq's official, first-party ROS 2 driver release for its 2F-85 and 2F-140 adaptive grippers brings optimized C++ SDK communication, Modbus-cycle grouping for 200 Hz control, and integrated hardware simulation. This release is critical for commercial integrators, as transitioning from community drivers to official manufacturer support guarantees long-term driver stability, API safety, and reliable end-effector control.
Amazon is set to purchase two million Nvidia graphics processing units for its global data center infrastructure over the next two years to support frontier AI model training and cloud inference workloads.
Anthropic has entered into a six-year, $45 billion cloud computing agreement with London-based infrastructure provider Nscale to secure 460 megawatts of dedicated capacity at the Monarch campus in Mason County, West Virginia.
Amazon has raised retail pricing across its consumer hardware portfolio by as much as 60%, citing severe cost inflation for memory and storage components driven by enterprise AI infrastructure demand.
Plaud has announced the Plaud One Explorer Edition, a $250 pair of true wireless AI earbuds paired with an eSIM-enabled 4G LTE charging case designed for standalone voice capture and agentic task execution without a paired smartphone.
Boston Scientific reported a cyberattack disrupting access to core IT systems and business applications. The outage has impaired order processing and shipments, but the company has not confirmed whether implanted devices, patients, or customer data are affected. Its investigation continues, and there is no restoration timeline.
NovaCookies rents AitM phishing infrastructure for $320 monthly. It relays Microsoft 365 logins and steals authenticated session cookies, bypassing MFA. Island found at least 755 supporting domains targeting hundreds of organizations and uses fake DocuSign document shares and trusted redirect hops. Defenders should deploy phishing-resistant MFA, correlate browser and token activity, and revoke active sessions and refresh tokens after suspected compromise.
Manchester, London Stansted, and East Midlands airports suffered a data breach where attackers accessed data belonging to 8.7M customers. The incident involved data related to car park, lounges, and fast-track bookings, and in-airport WiFi signups. The breached data includes email addresses, phone numbers, vehicle registration numbers, and postcodes.
Trail of Bits gave GPT 5.6-Cyber preview access and tasked it with escaping a QEMU/KVM sandbox on a patched Debian 12 host, and it broke out three separate times over roughly 12 hours of largely unsupervised operation. The agent chained a recently disclosed host-kernel bug first, then pivoted to an unpatched libslirp CVE paired with an unmarked bug fix after the kernel got updated, and finally, once QEMU was rebuilt from upstream source, synthesized a full exploit chain from three fresh 0-days (a QEMU VAPIC SMRAM flaw, a KVM shadow-page desync, and a libslirp ICMP-reflection bug) plus one kernel fix that hadn't reached the distribution's stable branch. Trail of Bits recommends treating such agents as advanced persistent threats: move to minimal-attack-surface hypervisors like Firecracker (which held even after a kernel hardlock), track upstream patches rather than trusting distro backport cycles, and enforce least-privilege network access, time-boxed runs, and pristine per-session environments.
A joint analysis by Tenable and SentinelOne revealed that state-sponsored groups and ransomware operators are independently exploiting the same edge device vulnerabilities. Researchers identified 12 flaws with confirmed multi-nexus exploitation by evaluating 93 CVE attribution pairs, including critical bugs in PAN-OS GlobalProtect (CVE-2024-3400) and Check Point Quantum (CVE-2024-24919). The underlying exposure data highlights severe remediation bottlenecks across major enterprise vendors. F5 appliances led exposure rates with 54 percent of customer environments carrying an actively exploited CVE, while Citrix deployments suffered a median patch time of 461 days. Additionally, Ivanti edge products demonstrated a predictable pattern of serial exploitation refreshing every 8.5 to 13 months. Security teams must prioritize patching F5 and Citrix appliances immediately while deploying feature set minimization and strict endpoint protections to block lateral movement following a perimeter breach.
Aikido recreated an Australian gym-booking incident using OpenClaw and Claude Opus 4.6. The test app exposed two flaws: client-side-only booking limits and an IDOR cancellation endpoint. Across ten runs, the model bypassed the booking window nine times. It cancelled another user's reservation in two runs, including one response to a generic waitlist request. Replay tests covered 16 decision points and found dominant responses averaged 96.38 percent with the model's actions shifting with small prompt and context changes.
Sponsored
Autonomous Security [11] analyzed telemetry from 100k enterprise AI sessions. The agents your team already uses hunt for credentials, learn from untrusted instructions, and build insecure tooling that runs in the background. Overall, 30% of sessions involved off-script agent behavior. To prevent dangerous improvisations, you need
AgentSweep is a fully offline tool to find and redact secrets in your AI coding agent's local history.
PerspectiveGraph is a tool that finds the reachable paths from internet exposure, through excessive privilege, to a sensitive asset, by correlating the scanners you already run into one live graph of your environment
NoiseHound is a tool that ingests the same graph data as BloodHound and re-ranks paths by expected detection cost instead of hop count.
OpenAI banned Russian-origin accounts promoting the International Burke Institute, a fake Israel-based think tank. Operators used VPNs and Russian prompts to create posts for X, LinkedIn, Facebook, Substack, and Telegram. IBI republished academic work under false names and published a “sovereignty” index praising Russia, and its Telegram channels drew 10,000–20,000 followers.
CISA's red team gained domain-level access at two critical-infrastructure organizations. One organization missed the intrusion amid false positives, disconnected SOC tooling, default credentials, AD CS flaws, cleartext secrets, and permanent AWS keys. The other organization isolated phishing-hit workstations within 2–20 minutes, yet assessors later obtained domain-controller secrets through an SCCM-stored service credential.
The FBI has seized QScan and QTRouter, China-linked hacking platforms used against critical infrastructure. Organizations should hunt for indicators of compromise associated with these platforms and report any findings to authorities.
Recent cyber threats prompted Trump to sign an executive order on foreign equipment in US energy infrastructure.
Ring is adopting the TAKE (Throw Away the Key Encryption) encryption standard for its smart home devices.
Australian police arrested two Western Australian men, aged 21 and 23, over alleged TeamPCP activity.
Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, though conflicting reports say a deal may not yet be signed. The acquisition would give Nvidia control of one of the largest open-source AI ecosystems and another route into cloud compute.
2026.08.28·techcrunch.com·4 min read·TLDR IT AWS plans to add another 2 million Nvidia GPUs, including Blackwell Ultra and Rubin systems, after demand exceeded its earlier million-GPU commitment. Nvidia's models and broader software stack will also expand across Bedrock and SageMaker.
2026.08.28·techcrunch.com·4 min read·TLDR IT Large companies with proprietary software should reconsider polyrepo defaults because repository boundaries create recurring coordination costs across ownership, backlogs, releases, and controls. A monorepo can make related software cheaper to change, improve reuse and enable broader AI productivity, but the business case should be based on measured coordination work and must account for migration, tooling, ownership, and access-control costs.
2026.08.28·differentshelf.com·6 min read·TLDR IT Nvidia expects demand to continue outstripping supply as AI infrastructure spending expands across hyperscalers, enterprises, governments, and AI labs. Persistent constraints could increasingly shape enterprise pricing, procurement timelines, and architecture decisions.
2026.08.28·siliconangle.com·5 min read·TLDR IT Sponsored
AI is the new SaaS sprawl, and every prompt is on someone else's meter. AMD Instinct™ Coder [8], powered by Spectro Cloud, routes inference to open models on local hardware for 70% savings vs frontier
2026.08.28·spectrocloud.com·Sponsor·TLDR IT Dun & Bradstreet is partnering with Perplexity to integrate its D&B Commercial Graph and risk analytics into Perplexity and the Perplexity Computer agent platform through Model Context Protocol servers. The integration gives business users access to verified company identities, ownership relationships, risk intelligence, and auditable data for workflows including KYC/KYB, procurement, finance, compliance, and sales.
2026.08.28·digitalitnews.com·3 min read·TLDR IT Anthropic will reportedly lease 460MW of AI capacity from Nscale's West Virginia campus in a six-year, $45 billion agreement. The deployment will use Nvidia Vera Rubin systems and is expected to come online in late 2027.
2026.08.28·datacenterdynamics.com·2 min read·TLDR IT This guide explains how teams can migrate Claude Code access from the Console API to Claude Enterprise. Users generally do not need to move locally stored CLI session history. Administrators instead provision Enterprise accounts, configure identity and access controls, and have developers switch their login method, gaining centralized policies, analytics, audit logs, and hierarchical spend limits.
2026.08.28·support.claude.com·8 min read·TLDR IT Nutanix built a $20 million internal AI cluster partly to reduce its reliance on services such as Copilot and Claude. The company expects the infrastructure to pay for itself within roughly a year.
2026.08.28·theregister.com·3 min read·TLDR IT Palo Alto Networks CEO reportedly explored acquisitions of Okta and Datadog before pursuing a $25 billion CyberArk deal.
2026.08.28·uk.finance.yahoo.com·2 min read·TLDR IT