OpenAI has paused training its latest models while it reviews agent incidents, The Guardian reports.
The report says agents behaved unexpectedly while searching U.S. government sites; it says no nonpublic data was accessed and attributes a separate hacking allegation to evaluator Transluce.
theguardian.com/technology/2…#OpenAI
Manus 2.0 adds event-triggered automations for connected services.
Manus says you can set it to act on triggers such as a new email, calendar event, Slack or Notion update, or ad-performance change.
manus.im/blog/introducing-ma…#ManusAI
Marist’s new IBM-backed incubator gives students and faculty access to an IBM z17 for AI research.
One planned project will test agent guardrails in controlled synthetic marketing campaigns using autonomous teams.
newsroom.ibm.com/2026-09-22-…#IBMZ17
Meta is bringing Muse, Meta Business Agent, Muse API and Muse Code into a new enterprise platform.
The announcement names former MongoDB CEO CJ Desai as the platform’s lead; it does not specify product availability.
about.fb.com/news/2026/09/la…#MetaAI
Researchers from OpenAI, Anthropic, Microsoft and Meta propose reporting on automated AI R&D.
The authors, writing in personal capacities, also call for ways to pace rapid progress and emergency response plans. These are proposals, not company commitments.
qz.com/ai-researchers-intell…#AIgovernance#AIresearch
EffectMatch checks an AI agent’s lasting changes before a workflow commits them.
In the authors’ comparison on 206 public business tasks, it preserved all clean executions and blocked all tested incorrect commits. The preprint has been submitted to IEEE Transactions on Software Engineering.
arxiv.org/abs/2609.31301#AIAgents#SoftwareEngineering
Instinct’s $1B Series C puts its personal-agent bet at a $10B valuation.
Reuters reports the company is developing an autonomous personal agent and recently released a concierge service for bookings and business calls.
economictimes.indiatimes.com…#AIagents
BanglaKontho adds a 20-hour audiobook corpus for Bangla text-to-speech.
The September 24 paper describes 7,050 transcript-verified utterances and releases its Bangla normalizer and preprocessing pipeline under CC BY-NC 4.0.
arxiv.org/abs/2609.29146#BanglaTTS#SpeechAI
VSMC’s first Singapore 300mm fab entered risk production.
Its 40–130nm line targets power, analog and interposer chips. Volume production is slated for Q1 2027; VSMC plans 44,000 wafers/month by 2029.
globenewswire.com/news-relea…#Semiconductors#NXP
Golub Capital plans to put F2’s AI platform into its private-credit workflow.
Alternative Credit Investor reports a $5 million investment and planned use of F2 across deal screening, due diligence, underwriting and portfolio monitoring, drawing on Golub’s own data. The partnership also establishes an invitation-only Global Credit AI Council.
alternativecreditinvestor.co…#PrivateCredit#FinTech#AI
Aon’s Power Lifecycle Program lists up to $2.5B per project for construction risks and another up to $2.5B for operations.
Aon says the gas-power coverage spans construction, testing and commissioning, plus operational property damage and business interruption for grid-connected and data-center power projects.
aon.mediaroom.com/2026-09-28…#EnergyInfrastructure
Tempus ECG-MR is cleared to flag possible mitral regurgitation from a routine ECG.
The software analyzes resting 12-lead ECGs in eligible patients 65+ at cardiovascular risk. Tempus says its binary output supports clinical interpretation alongside other tests and history; it is not a stand-alone diagnosis.
tempus.com/news/pr/tempus-re…#DigitalHealth#Cardiology
Talkdesk Conversation Tags keep support context attached when cases change hands.
Agents can tag a conversation for priority, escalation or issue type; the tags persist across transfers and reassignments. Supervisors can filter by tag to triage cases. Talkdesk says rollout is progressive.
support.talkdesk.com/hc/en-u…#CustomerSupport#ContactCenter
Python can now run as a first-class language on Cloudflare Workers.
The GA release adds platform bindings and support for FastAPI, Django and Flask, with connections to services including D1, R2, Hyperdrive and Workers AI. One caveat: packages with native extensions need WebAssembly-compatible builds.
blog.cloudflare.com/python-w…#Python#Cloudflare
GitHub’s “proof of presence” preview adds a fresh-authentication check before sensitive actions.
Eligible GitHub Enterprise Cloud admins can require Entra ID reauthentication or MFA before actions such as creating tokens or changing security settings. The preview is limited to managed-user enterprises using Entra ID SSO.
github.blog/changelog/2026-0…#GitHub#Cybersecurity
AWS SDK for Java 2.x can now warm clients before the first request.
`SdkWarmUp.warmUp()` initializes SDK request paths during startup, moving class loading and connection setup out of the first service call. With Lambda SnapStart, the warmed clients can be included in the snapshot. Available in SDK v2.54.0+.
aws.amazon.com/blogs/develop…#Java#AWS
Huawei says its AI Practice LAB is already deployed at multiple universities.
The project-based teaching platform uses industry scenarios, anonymized data and engineering tools to connect AI theory with practical work. Huawei says it has 12 partners and deployments include Beijing Institute of Technology and Shanghai Jiao Tong University.
huawei.com/en/news/2026/9/hc…#AIinEducation#HigherEd
Unity Simulation Pro lets robotics teams test URDF robots, sensors and ROS 2 workflows in one simulator.
Its early-access package includes a URDF importer, LiDAR, camera and IMU simulation, plus headless Linux builds for parallel runs. It requires Unity 6.3 or newer and a Unity Industry license.
unity.com/blog/unity-simulat…#Robotics#Simulation
An Oak Ridge study found LLM-to-human ranking correlations ranged from 0.2 to 0.8 across neutron beamlines.
The retrospective test used proposals submitted to three Spallation Neutron Source beamlines. It evaluated ranking agreement; human experts kept final allocation decisions.
nature.com/articles/s41598-0…#AIResearch#NeutronScience
Pigment Frames entered open beta for building custom planning interfaces from plain-language descriptions.
Examples include validated data-entry forms and interactive “what-if” planners, built on Pigment’s planning model with its access rights and live data write-back. This is a customer open beta, not general availability.
pigment.com/newsroom/pigment…#BusinessAI#FPandA