AI News for August 29, 2026
Browse any day on a calendar, install it to your home screen, read offline.Read in the app →-
1TechCrunch20:46
Anthropic gets its first court win over the Pentagon’s supply-chain risk label
A federal judge ruled that the Trump administration illegally designated Anthropic a supply-chain risk, a label that had effectively shut the company out of Pentagon procurement. It is Anthropic's first legal win in the dispute, and a second lawsuit against the Pentagon continues in Washington. Ars Technica also covered the ruling, noting the conflict began when Anthropic refused to support lethal autonomous warfare and mass surveillance uses.
Read original ↗ -
2VentureBeat00:33
Meta researchers taught an 8B AI model to match Claude Opus 4.5 — without the frontier price tag
Meta researchers report that an 8-billion-parameter model trained to exploit execution feedback from the runtime harness can match Claude Opus 4.5 on long-running agent workflows. The target is enterprise tasks that far exceed a single context window — such as multi-hour record migrations between a legacy CRM and a cloud database — where the agent must ground itself in server logs and other environment signals rather than its own memory.
Read original ↗ -
3TechCrunch03:30
An Anthropic researcher just gave us a peek at self-improving AI
An Anthropic researcher shared early results from automated systems that improve a model's own behavior. Given 10 benchmarks covering specific misaligned behaviors, the systems raised performance on every one without degrading overall capability. The work applies self-improvement loops to alignment rather than raw capability, an area where automated iteration has been harder to demonstrate.
Read original ↗ -
4TechCrunch04:24
Neocloud Lambda secures $1B in debt to buy more chips
GPU cloud provider Lambda raised $1B in private debt to buy Nvidia AI chips and lease the capacity to Microsoft. It is the latest in a string of loans taken on by neoclouds, which are financing hardware purchases with debt rather than equity and carrying the cost of the buildout on their balance sheets.
Read original ↗ -
5VentureBeat23:30
Cohere Parse 5 loses the benchmark on points. It wins on cost per page.
Cohere released Parse 5, a 2.3-billion-parameter vision language model for turning PDFs, slides and scanned documents into structured input for AI pipelines. The company is positioning it on price-to-performance rather than top benchmark accuracy, arguing that enterprises running documents at scale are constrained by cost per page and by tools that lose tables, charts and layout.
Read original ↗ -
6TechCrunch02:19
Open-weight AI companies are the Valley's hottest acquisition targets
TechCrunch reports that companies releasing open-weight models have become prime acquisition targets, with substantial capital flowing into businesses whose core product is given away. The trend reflects buyers valuing the teams, distribution and ecosystem position that open releases create, even without direct model revenue.
Read original ↗ -
7The Verge00:28
Trump’s EPA wants to let data centers hide their air pollution
The US Environmental Protection Agency plans to scrap a federal rule that requires public notice and a comment period when certain industrial sites seek air permits. The change would make it harder for residents to contest pollution from new data centers, which are already facing organized local opposition as AI buildouts expand.
Read original ↗ -
8TechCrunch20:21
Meta executive leaves for OpenAI as the social media giant faces growing scrutiny in India
Sandhya Devanathan is leaving Meta for OpenAI, where she will oversee some operations across Southeast Asia and Australia. The move comes as Meta faces increasing regulatory scrutiny in India, and adds to OpenAI's push to build out regional commercial leadership in Asia-Pacific.
Read original ↗ -
9Ars Technica23:42
Meta makes AI glasses slightly less creepy with limit on nonconsensual recording
Meta updated its AI glasses so they stop recording when the safety light that signals capture is covered up, closing a workaround that allowed silent, nonconsensual filming. Ars Technica notes the fix narrows one abuse path but leaves broader privacy concerns around always-available camera wearables unresolved.
Read original ↗ -
10VentureBeat01:52
The three layers of agentic AI security: A defense-in-depth architecture for autonomous agents
In a piece presented by Nutanix, product management director Oscar Wahlberg argues that autonomous agents that reason, decide and act across an environment create risks application-level controls were never designed to hold. He proposes separating the problem into three layers, on the grounds that guardrails built to catch a malicious prompt do not address failures at the infrastructure and execution layers.
Read original ↗
