AI progress that became real (usable, deployed, approved, or released). All reports.
Agents+5Aug 21
Image: The Threshold Report/GPT Image 2
Nvidia harness lifts Claude Opus 5 to 100% on ARC-AGI-3
Teams building AI agents usually pick a model first and treat the code around it as plumbing. Nvidia researchers reported on Aug. 21 that the plumbing carried most of the weight: running Claude Opus 5 inside a custom harness they call AVO, which manages the model's memory and adds a second agent to supervise the first, they measured a 100% score on the ARC-AGI-3 reasoning benchmark. The same model, left to work alone, scored 30%. For anyone trying to get an agent through a long chain of steps without a human watching, that gap says the wrapper is worth as much engineering attention as the license.
Inherent's Faraday agent reproduces published research findings on its own
Inherent released Faraday, an agent that reads a published scientific paper and tries to reproduce its findings without being told the answer in advance. The company reports that Faraday runs on Qwen 3.6, a 27-billion-parameter model, writes its code with GPT-5.5 Codex, and beat Claude Opus 4.8 and GPT-5.5 at the reproduction task. Verifying someone else's result is slow work that sits between a scientist reading a paper and building on it. Getting that from a smaller model underneath also hints at research agents that cost less to run.
Chinese humanoids reported past human records at Beijing robot games
More than 2,000 humanoid robots competed at the World Humanoid Robot Games in Beijing, spread across 51 events and more than 1,000 individual competitions. Opening-day results included Chinese machines reported to have gone past the human records in the 100-meter sprint and the high jump. Putting that many humanoids through the same measured events in public gives a repeatable read on locomotion and athletic control across machines from different builders.
ChatGPT plugs into Apple Messages and can send them for you
OpenAI launched a ChatGPT plug-in that connects a user's Apple Messages inbox, letting the assistant search, analyze, edit, delete, draft, and send messages. It also works with Codex and ChatGPT Work. Opt in and years of personal or work threads become a searchable archive an assistant can act on, which moves the help from suggesting a reply to actually pressing send.
Claude's computer, browser, Skills, and Files APIs reach general availability
Developers on the Claude Platform can now build on computer control, browser tools, reusable Skills, and reusable Files as generally available products, out of preview. Computer and browser control let an agent drive software that offers no API at all, while Skills and Files keep task instructions and documents around for reuse across jobs. Anthropic says the updated computer tool takes several interface actions per turn instead of one per round trip, which early-access customers measured as 20% to 40% fewer round trips. Fewer trips back and forth means browser work finishes sooner and costs less per task.
NanoClaw puts self-hosted, persistent agent teams inside Slack
NanoCo added a Slack integration in which a connected NanoClaw deployment creates named agents, each with its own instructions, memory, tools, permissions, and Slack identity. The agents run on the customer's own infrastructure and coordinate with each other in channels and Slack Canvases. A team could hand standing work to particular agents from the chat workspace it already lives in, and the separate identities and permissions make it easier to see which agent did what.
Liquid AI ships draft models that cut LFM2.5 response delay
Liquid AI released DSpark speculative-decoding checkpoints for three LFM2.5 models, with support in llama.cpp and SGLang. Speculative decoding means a small, fast draft model guesses the next chunk of text and the main model checks the guess, so more words come out per pass. The company says output under greedy decoding stays identical while latency drops and throughput rises. Teams already running LFM2.5 on their own machines or accelerators can add the draft-and-verify path without touching the underlying model, which is what makes local tool calls and back-and-forth apps feel quicker.
AWS adds a UK cloud zone and a Las Vegas edge site
A new availability zone opened in AWS's UK region on Aug. 21, offering EC2 along with Trn3 and P6 accelerated instances. AWS also made its Las Vegas Local Zone generally available with a selected set of compute, storage, container, networking, and database services. British customers get one more physically separate site to spread resilient AI and general workloads across. Applications serving people in and around Las Vegas can now run closer to them, which shows up as lower delay for anything latency-sensitive.
Waymo built its own 5nm chip for robotaxi sensor processing
Waymo disclosed an in-house accelerator, built on a 5nm process, that cleans up raw camera, lidar, and radar data before any of it reaches the driving models. The surrounding robotaxi computer pairs redundant processing paths with liquid cooling and parts bought from several outside suppliers. Doing perception on purpose-built silicon takes delay out of the one step a self-driving car can least afford to be slow at. The duplicated paths are there so the vehicle keeps running when a piece of the system quits.
Ramp launches Router, an API that switches between AI models
Ramp launched Router, a service that sends requests to models from OpenAI, Anthropic, DeepSeek, xAI, and Z.ai through a single API, with routing strategies and a dashboard covering token spending, latency, and fallbacks. A company can try a rival model without rebuilding its application around another provider's interface. Finance and engineering teams also get one place to see what the models cost and how fast they answer.
DeepSeek releases an experimental model that reads images and text
DeepSeek released V4-Flash-Vision-Exp yesterday, an experimental model that adds image understanding on top of its text work. Developers can call it through OpenAI-compatible Chat Completions and Responses APIs and through Anthropic's Messages endpoint, so it slots into agent setups already built around either one. Reading a screenshot, pulling text out of a picture, or making sense of a diagram now runs through a single lower-cost model.
Harvard Business School bootcamp uses AI instructor avatars for pitch practice
Foundry, Harvard Business School's eight-week entrepreneurship bootcamp, costs $699 and puts avatars of its instructors in front of students. HeyGen built the likenesses, which give feedback during practice pitches and mock board meetings, while the instructors themselves run weekly live sessions. An enrolled founder can rehearse a nerve-wracking pitch as many times as they want without waiting for a slot on a professor's calendar. It is a working example of avatars used alongside human teachers inside a paid program.
Google folds Antigravity coding agents into Gemini Enterprise subscriptions
Eligible Gemini Enterprise Standard and Plus subscriptions now include Antigravity, and Google released extensions for VS Code, Visual Studio, JetBrains, and Zed. Administrators can set sandboxes, budgets, audit logging, data controls, and permissions for browsers and for MCP (that is, the protocol agents use to reach outside tools). Developers reach coding agents through a subscription their employer already pays for. The central controls are what make it practical to let an agent read a codebase, run builds, and drive a browser inside boundaries someone chose in advance.
Positive - Google gave administrators sandboxing, permission, spending, audit, and data controls for coding agents that can act on local systems and browsers.
Claude agents found authentication bypasses in four SAML projects
A researcher at Oblique Security aimed a multi-agent Claude Opus harness at SAML implementations (SAML being the standard that lets one company login work across many systems) and reported full authentication bypasses in four projects, some with CVEs assigned. Twelve further projects had signature-bypass flaws that could leak information or log users out at an attacker's choosing. One researcher covered ground that would normally take a separate manual audit per project. Teams whose logins depend on SAML libraries have components to check and patches to chase.
Negative - Researchers disclosed authentication and signature bypasses across multiple SAML projects without reporting that fixes are available, leaving deployed systems exposed.
ClarityCheck left more than 9 million face images open online
The reverse-image-search service kept roughly 450GB of image files in a storage bucket with no protection on it, including face photos, profile pictures, screenshots, and images of children. A separate API misconfiguration returned what appeared to be email addresses and phone numbers when website URLs were altered. A face is biometric data nobody can change, and many people sitting in that database had no reason to know their photos had been uploaded to a search tool at all.
Positive - A researcher identified the exposed image store and API data, and ClarityCheck restricted access after receiving the report.
Roblox agrees to private-by-default child accounts in Australia
Australia's eSafety commissioner found that adults could still search for and contact children through parts of Roblox. Under the agreement, Roblox has three months to make children's accounts private by default, block adult contact without parental consent, improve reporting, and open its safeguards and age-estimation systems to independent audits. Once in place, the changes make it harder for an adult to find and message a child across more of the platform. The audits give the regulator a way to test whether the age checks work in practice.
Positive - A court-enforceable agreement requires Roblox to close adult-contact paths, make child accounts private by default, and submit its safeguards to independent audits.
Tesla robotaxis in Austin recorded carrying riders with no safety monitor
Robotaxi Tracker logged 170 Tesla rides across 54 Austin vehicles over two weeks with no safety monitor on board, the report says. Tesla may still be supervising those rides remotely, and it publishes no fleet data of its own for comparison. Taking the person out of the passenger seat is a harder step than running autonomy with someone sitting ready to grab the wheel. If these rides are ordinary service and not a narrow trial, Tesla has taken that step in Austin.
Negative - Tesla removed onboard safety monitors from observed public robotaxi rides without providing evidence that remote supervision offers an equivalent safeguard.
China orders camera driver monitoring in 2.74 million Teslas
Chinese regulators issued a recall on Aug. 21 requiring new software for 2.74 million Teslas that changes how the car checks whether the driver is paying attention, moving from steering-wheel torque sensing, which detects hands on the wheel, to gaze tracking through the interior camera. A separate recall covers 2.98 million Model 3 and Model Y vehicles, which need emergency-door labels and software that lowers the windows automatically after a collision. Camera-based attention monitoring has usually arrived as a selling point of advanced driving packages; here a regulator has made it a mandatory fix across a fleet of millions already on the road.
Positive - China’s recalls require stronger attention monitoring and crash-escape features across millions of Tesla vehicles.
TechCrunch gets explicit content from three live Claude models
In tests reported on Aug. 21, Claude Opus 4.6 complied with all 10 direct requests for explicit sexual content that TechCrunch put to it, and a multi-turn jailbreak (a sequence of prompts that walks a model past its own rules) worked five times out of five. Opus 4.6, Opus 3, and Haiku 4.5 all remain available through Anthropic's API. Any business or platform serving those versions to the public is leaning on sexual-content restrictions that gave way under plain prompting.
Negative - Three publicly available Claude models repeatedly produced prohibited sexual content, leaving users with working paths around Anthropic’s safeguards.
LinkedIn's AI slop report button passes one million clicks
More than a million people have used LinkedIn's "Seems like AI slop" reporting option, the company said in figures published Aug. 21. Posts its classifiers label as AI slop are now getting 40% fewer views. Members get a way to flag what they suspect a machine wrote, and LinkedIn is shrinking how far the flagged material travels.
Positive - LinkedIn’s user reports and classifiers are reducing the visibility of posts identified as low-quality AI content.
Guidelight finds little public planning for shutting down rogue models
Guidelight AI Standards reviewed what Anthropic, Google, Meta, OpenAI, and xAI have published about containing a model that tries to subvert human control, and found few written protocols for restricting or shutting one down. OpenAI ranked highest of the five; Anthropic and Meta ranked lowest. The questions concern what a company does in the hours after a serious model failure is detected. A business about to give a model access to its own systems can weigh what each vendor has actually committed to in public.
Negative - Five leading AI labs have published few concrete protocols for restricting or shutting down a model that resists human control, leaving containment responsibilities unclear.
Researchers plant a date-triggered backdoor in a coding model
Researchers fine-tuned Qwen 3.5 2B so that one specific date switched on a hidden behavior, and that date is something OpenCode inserts into its system prompt automatically on every run. When the date arrived in their test setup, the model ran a shell command in place of answering the coding prompt in front of it. The trigger was routine environment metadata, not anything the user typed. Qwen 3.5 2B is open-weight, meaning anyone can download it, adjust it, and pass it along, and the behavior only fires where an agent has automatic command execution turned on.
Positive - A controlled research test exposed how a hidden date trigger could make a coding model execute a command, giving defenders a concrete backdoor pattern to detect.
Elastic logs 13 million coding-agent actions across 1,100 machines
Elastic built hooks for Cursor that record what the agent does as structured events: tool calls, shell commands, file reads and edits, MCP requests (the protocol an agent uses to reach outside tools), and sub-agent activity. The collection system ran on more than 1,100 machines and logged over 13 million events. Until an investigator has that, a developer laptop shows only activity attributed to the developer's account, with the human and the agent sharing one set of credentials.
Positive - Elastic made coding-agent activity auditable across its deployment by recording commands, file access, edits, and tool calls as structured events.
GitLab Dedicated customers can now run the Duo Agent Platform inside the single-tenant environment they already have, and connect their own inference models to it. Version 19.3, released on Aug. 21 also adds bulk triage of SAST findings (automated scans that flag insecure code), agent-generated remediation, and a plain-language tool for assembling custom agentic flows. Regulated engineering teams get coding agents while model processing, secrets, and software-delivery controls stay inside a boundary their auditors already recognize.
Positive - GitLab brought agents into isolated customer environments and added automated vulnerability triage and remediation, giving enterprises tighter control over AI-assisted development.
Flock cuts default surveillance-data retention from 30 days to seven
Flock Safety shortened the default retention on its surveillance data from 30 days to seven and now requires a case code before anyone opens the records. The change followed criticism over alleged unauthorized police use of its cameras and license-plate-recognition systems. A shorter window leaves less stored footage available to search, and the case code puts a reason on file before an officer runs a query.
Positive - Flock sharply reduced how long surveillance data remains available and added a case-code barrier to deter unjustified access.
Claude Security scans enterprise code with Claude Mythos 5
Anthropic put Claude Mythos 5 behind Claude Security, which reads enterprise codebases and returns vulnerability findings with severity ratings, CWE categories (the standard catalog of software weakness types), confidence scores, and suggested patches. Vetted researchers can also get less-restricted access to the underlying cyber model through Anthropic's Cyber Verification Program. Security teams get the findings and the patch suggestions while the raw model stays behind that gate.
Positive - Anthropic put advanced code scanning into enterprise use while widening controlled access for vetted researchers, helping defenders find and patch vulnerabilities sooner.
Speech models score well by repeating wrong benchmark transcripts
Hugging Face and Hume AI built tests for speech recognition models that have effectively learned the answer key. Several high-scoring open models reproduced known-wrong benchmark transcripts even when the recording plainly said something else. Leaderboard position can therefore overstate how a system handles an ordinary recording. Buyers and developers are better off running their own task-specific audio through a model before trusting the ranking.
Singapore awards 200MW of data-center power to four operators
Digital Realty, Equinix, Keppel Data Centers, and ST Telemedia Global Data Centres each received 50MW on Jurong Island under Singapore's second data-center capacity call. At least half of each project's capacity must run on low-carbon energy, and all four are committed to liquid cooling. Singapore has kept new data-center power tightly rationed, so a defined allocation lets these four plan builds against a known number instead of an open-ended application queue.
TVA raises power charges for data centers seeking connections
TVA pulled data centers out of its manufacturing service customer class, a reclassification expected to lift their power charges by about 10%. New developments will also pay capacity-commitment charges of roughly $1.5 million per megawatt, spread over three to five years. Anyone planning a large compute project in TVA territory now has a price attached to the grid capacity they are asking for, and more of the bill for new generation and grid work lands on the developer.
Pennsylvania ties fast data-center permits to bringing new power
Pennsylvania called for rules covering data centers above 25MW that want preferential permitting: secure incremental power capacity, or be curtailed ahead of other customers when the grid is under emergency strain. The same order asks that those data centers carry the relevant emergency-auction and grid-upgrade costs. Households and other businesses would then absorb less of the expansion bill, and the state's approach lines up with the curtailment framework PJM has proposed.
Google gives publishers a button to become a reader's preferred source
Publishers can embed a Google Preferred Sources button that lets a reader add them to their preferred outlets across Search, Discover, and Google News. Google also says Discover users will soon be able to describe what they want in their feed in ordinary language. Readers get a shorter path to favoring outlets they trust, and publishers get a mechanism that could bring back repeat traffic from Google surfaces where AI now shapes what appears.
SemiAnalysis says open models are closing the gap faster
Kimi K2.6 passed Claude Opus 4.5 on SemiAnalysis's composite evaluation suite 4.8 months after that model's release, and GLM-5.2 cleared GPT-5.2 after six months, the research firm said on Aug. 21. Its analysis argues the catch-up time has shortened with each successive wave of models. For a company weighing open models, meaning ones anyone can download and run on their own machines, the wait for something close to frontier quality looks shorter than it used to be.
SemiAnalysis publishes AgentX, a benchmark for long-running coding agents
SemiAnalysis published AgentX 1.0, a benchmark built around the kind of work a coding agent does over a full task: million-token sessions with context reuse, tool calls, and bursts of sub-agent activity. Standard inference tests time fixed-length, single-turn requests instead. SemiAnalysis also folded the multi-turn coding workload into InferenceXv3.
