1. 100+ AI and Security Companies Warn of a “Limited Window” on AI Cyberattacks
On August 27, OpenAI, Anthropic, Google, Microsoft and more than 100 other signatories published a joint open letter warning that AI-enabled cyber attacks will become “far more widespread and sophisticated” within months as models grow more capable. Signatories included CrowdStrike, Okta and Fortinet alongside financial institutions and internet infrastructure firms. The letter argues there is only a months-long window to harden defences for hospitals, water treatment plants and core internet infrastructure, and calls for coordination at local, national and international levels.
2. Nvidia in Talks to Back Perplexity at a $30B+ Valuation
Nvidia is reportedly in discussions to join a new Perplexity financing round valuing the search startup above $30 billion — more than 50% above the roughly $20 billion mark set a year ago. Perplexity’s annualised revenue has climbed past $750 million, up from under $250 million at the start of 2026, growth partly credited to Perplexity Computer, its cloud-based agent for automating desktop work. Nvidia has held a stake since 2023; the company is targeting an IPO around 2028.
3. The EU AI Act Gets Real: Enforcement Powers Now Live
The August 2 compliance deadline has passed and the EU AI Office now holds live enforcement powers over general-purpose AI models — it can demand technical documentation, evaluate models, order corrective measures and levy fines. Transparency and disclosure duties for chatbots and AI-generated content are in force. High-risk obligations for stand-alone systems remain deferred to December 2027, and to August 2028 for AI embedded in already-regulated products. Vendors now face overlapping EU, US state (Colorado, California) and US federal regimes simultaneously.
4. Anthropic Ships a Built-In Browser Inside Claude Cowork
Starting August 26, Anthropic’s desktop agent app Cowork includes its own Chromium browser in a side panel, so Claude can navigate, read pages, click and fill forms without driving the user’s own browser. It runs on a clean sandboxed profile with no stored logins and requires no extension or setup. The feature is live on Enterprise and rolling out to Pro, Max and Team users on macOS, Windows and Linux.
5. Google Opens Managed Agents and Computer Use in the Gemini API
Google pushed a run of agent-focused Gemini API updates in late August: Managed Agents entered public preview, letting developers deploy autonomous, stateful agents inside secure Google-hosted Linux sandboxes. The Computer Use tool also went to public preview on Gemini 3.5 Flash with browser, mobile and desktop support plus configurable safety policies. Separately, gemini-3.1-flash-lite-image (Nano Banana 2 Lite) reached general availability for low-latency image generation and editing.
6. GEN-1.5 Claims One-Shot Robot Learning From a 3–12 Second Demo
Generalist AI released GEN-1.5, an embodied foundation model that picks up new physical tasks from a single 3–12 second demonstration with no retraining or coding. It consumes video, sensor, language and proprioceptive input and emits action trajectories at 100 Hz, holding 30 seconds of video context after more than eight months of pretraining on physical interaction data. Across 10 tasks — opening jars, unzipping a pouch, retrieving money from a purse, sweeping — it averaged 59% success, rising to 83% with about five minutes of data and 10 training steps. The company raised $400M in June at a $2B post-money valuation.
7. New Numbers Land for OpenAI’s Jalapeño Inference Chip
Fresh performance detail emerged this month for Jalapeño, OpenAI’s first custom inference ASIC, built with Broadcom on TSMC’s N3P process and taped out in roughly 16 months. Benchmarks on models such as GPT-OSS 120B show 1.5–1.9× more tokens per kilowatt and up to 3.6× lower latency than Nvidia’s GB200 and GB300, with larger gains on interactive workloads. The 700 W part carries 216 GB of HBM4 at 15.4 TB/s and posts 700+ tok/s/user on DeepSeek R1. Deployment is targeted for late 2026.
8. Amazon Is Shutting Down Mechanical Turk After 21 Years
Amazon announced on August 25 that Mechanical Turk closes for good on September 30, 2026, taking SageMaker Ground Truth with it. Launched in 2005 and once serving more than 500,000 workers on data labeling, transcription and survey tasks, the platform had been in decline as Scale AI, Mercor and Prolific pulled workers away. Amazon had already stopped accepting new customers a month earlier. Jeff Bezos once described the service as “artificial artificial intelligence” — the human labour layer that helped bootstrap the modern AI industry.
// KEY TAKEAWAYS
Agents are the through-line this week: Claude got its own browser, Gemini got managed agents and computer use, and GEN-1.5 pushed the same in-context-learning trick into robotics. The flip side arrived on the same calendar — 100+ companies warning that capable models cut both ways on cyber defence, and the EU AI Office switching on real enforcement powers. Meanwhile the money and the silicon keep moving: Nvidia circling Perplexity at $30B, OpenAI’s own inference chip posting real perf-per-watt numbers against Nvidia parts, and Mechanical Turk — the human-labour scaffolding the whole field was built on — quietly closing down.