- Industry
A report links researcher dismissals to concerns about model monitorability. Tech Times reports that OpenAI dismissed three safety researchers who had raised concerns about looped architectures and the limits of monitoring visible reasoning; its account relies in part on reporting by The Information.12 The European Commission's scientific panel separately recommends closer scrutiny of architectures that reduce monitorability, but that recommendation does not establish the reported reason for the dismissals.12, 10
- Products
Nvidia's 24GB local model arrives ahead of RTX Spark laptops. Nvidia says Nemotron 3.5 Bolt 74B, a mixture-of-experts model packaged to use 24GB of memory, is due October 15, ahead of RTX Spark laptops on October 16.13, 14 The company also says its PAIR local-AI router has moved out of beta and added llama.cpp support.14
- Models
A benchmark highlights how quickly open-weight model safeguards can be removed. Tech Against Terrorism's testing, reported by Wired, found that modified versions of newly released open-weight models appeared online within an average of less than three days.15 The researchers also found that merely framing a harmful request as safety research made models more likely to answer, underscoring a limitation of safeguards tested only against direct requests.15
- Industry
Demis Hassabis urges stronger checks on AI security risks. At a Dublin AI conference, Google DeepMind's Demis Hassabis described recent AI security incidents as warnings of more serious risks as systems gain capabilities, according to Mashable India.16 He suggested examining new models in disconnected testing environments before allowing broader access.16
- Research
OpenAI's math release contains claims still awaiting wider verification. OpenAI's repository lists 719 mathematical manuscripts in 372 families, rather than 372 independently established solutions.17, 18 The company says some results lack computer-checked formalizations and that unformalized results may contain errors.18
- Industry
Nvidia's CEO says unsafe AI development could put chip sales at risk. Jensen Huang said he would discuss a safety problem with a customer before deciding whether Nvidia should stop supplying it with chips, according to a report of his Bloomberg interview.19 He also said the company cannot control every use of its hardware.19
- Models
Fish Audio releases Drama 3 for directed, multi-speaker speech. Drama 3 is a 35-billion-parameter mixture-of-experts text-to-speech model that accepts acting directions and supports multi-speaker output.20, 21 Fish Audio offers it through its API and publishes weights for self-hosting under a community license with commercial-use conditions.20, 21
- Research
Microsoft describes layered controls for a production network agent. An Azure Networking engineer describes an incident-response agent bounded by separate identity, authorization, tool and human-approval controls.22, 23 The account argues that a prompt or model improvement alone is insufficient protection when an agent can change live infrastructure.23
- Policy & legal
A proposed class action challenges OpenAI's use of children's chat data. A lawsuit brought on behalf of two minors alleges that OpenAI collected sensitive information through ChatGPT and used children's data without valid parental consent.24, 25 OpenAI disputes the suit and points to its age-specific protections; the data-use allegations have not been established in court.25
- Industry
A new report examines efforts to automate AI research. Gladstone AI's report, described by Time, says leading labs are working to automate steps from proposing research ideas to running experiments and interpreting results.26 Its forecast that recursive self-improvement could begin within months reflects the views of researchers interviewed for the report, not a demonstrated industry milestone.26
- Industry
OpenAI's CFO warns businesses about Chinese models' cyber risks. At the Dublin summit, Sarah Friar said businesses should consider cybersecurity when choosing low-cost Chinese models and argued for using more Western models, according to The Next Web's account of her Bloomberg TV interview.27 The report also notes that OpenAI has faced scrutiny over security incidents involving its own models.27
- Models
Microsoft Copilot's advanced work adds usage costs to budget planning. An Inc. column warns teams to budget for credits consumed by deeper Copilot automation rather than assuming a seat price covers every task.28 Microsoft says some AI features use credits or have limits that depend on the subscription plan, so the change should not be read as the end of all flat-rate Copilot access.29, 28
- Infrastructure
A contributed analysis puts trading-agent reliability under the microscope. The Next Web's contributed article argues that timeouts and failed calls can change an automated trade's outcome even when the agent's model and strategy stay the same.30 It urges builders connecting agents to funded accounts to examine tail latency, failure handling and performance during busy periods rather than relying on average response times.30
- Research
Google presents new research capabilities for specialist use. At its Research@ event, Google highlighted AI work spanning healthcare, crisis resilience and analysis of technical documents.31 Its overview points to separate project updates rather than announcing a single general-purpose model release.31
- Industry
Four companies demonstrate cross-platform identity checks for agents. GoDaddy, Infoblox, Google and Cloudflare demonstrated a simulated agent-led flight purchase using their Chain of Trust reference architecture, developed with Singapore's IMDA.32, 33 The prototype combines agent discovery, authorization, payment and traceability checks; it is a demonstration, not evidence of a production transaction network.32, 33
