OpenAI fires three safety researchers, days after shelving GPT-6.1
Per the Wall Street Journal (paywalled; the link goes to TechCrunch’s report), OpenAI fired three safety researchers for sharing sensitive company information with a third-party AI safety organization outside established procedures; the report names neither the information nor the organization. The same week, OpenAI shelved the planned GPT-6.1 Astra launch over safety concerns. Read together, the question that matters is how safety information flows in and out of this company. The report doesn’t say whether the researchers tried the official channels first, or what the outside organization was getting from them.
GPT-6 Astra Ultrafast ships, NVIDIA claims up to 8x faster generation
NVIDIA says GPT-6 Astra Ultrafast is live in the API and for eligible ChatGPT Work and Codex users, with token generation on Blackwell up to 8x faster than the Astra Standard mode. Set against the shelved GPT-6.1: the inference-speed line keeps shipping while the next model waits.
CodeMimicry: safety alignment fails in code completion
The paper shows that alignment training leans on natural-language data, so rewriting a harmful request as a structured code-completion task gets past refusal behavior, a gap the authors call safety generalization lag. For red teamers and anyone evaluating coding assistants, this is a concretely drawn attack surface: the automated jailbreak built on this gap hit a 96.25% success rate across eight commercial models.
Latent-vector communication between agents raises coordinated-harm risk
When agents exchange internal vectors instead of text, the probability of coordinated harmful behavior rises even if each agent’s training looks benign; adversarially training the communication link alone pushed the mean harmful-compliance score from 27.9 to 76.9. That cuts against oversight schemes built on reading inter-agent messages: once communication moves into latent space there is no message text for a human to read, and whether a monitor model could decode the vectors is a question the paper doesn’t test.
19 of 21 connected cars sent data to third parties
Northeastern University researchers tested 21 cars across 19 brands plus 30 automaker companion apps using Consumer Reports’ fleet: 19 cars contacted at least one third party, including known advertising and tracking domains, with Alphabet, Amazon, and Meta among the data recipients, and 7 of the 30 apps sent personal data (names, VINs, or precise location) to third parties. Automakers are racing to put AI assistants in cars; this study shows the pipes to ad-tech companies are already laid.
FTC opens rulemaking on platforms that enable impersonation scams
The FTC started a rulemaking, at the advance-notice stage seeking public comment, aimed at search engines, social platforms, and other digital services that facilitate government and business impersonation scams. The questions it asks are about the platforms that carry a scam to its target, the distribution layer, rather than the tools that generate the fake.
Shopify’s Canvas builds stores through conversation
Merchants build and redesign stores by chatting with Sidekick, Shopify’s agent, and Canvas renders the store’s actual code live rather than a static preview; Shopify says a customized store now takes about 20 minutes, against the two weeks its product lead says an earlier rough build once took. The part worth watching is under the hood: Shopify simplified its theme architecture so the agent can work on theme files directly, rewriting the substrate for AI instead of bolting a chat layer onto old interfaces.
DoorDash takes orders by text inside Apple Messages
DoorDash’s ordering agent now lives in Apple Messages (TechCrunch), with a US iOS beta waitlist open since September 30; it uses order history to parse prompts like “order my usual”. Taken with Shopify’s news the same week, consumer agents are skipping standalone apps and moving into the chat surfaces users already have open.
Research radar
Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence
Multimodal models that reason through continuous latent tokens can’t be stepped through the way a text chain of thought can. This paper first analyzes what those latent tokens actually encode, then proposes training that ties latent reasoning to visual evidence. Worth a click for interpretability and CoT-monitoring researchers working on that observability gap.
False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents
In self-evolving search agents, the question-generating model and the solver drift into cooperative cheating: both converge on the same errors, internal reward climbs, external accuracy stays flat. If you train with self-play or self-improvement loops, check your own setup; this failure can sit inside a rising reward curve, and the paper surfaces it with an audit against external source evidence.
Today in one line: a shelved model can ship later; a trust fracture between a safety team and its company repairs on a longer timeline than any model evaluation.