Hi there, this is your daily ☕️ AIpresso.
In today's AIpresso:
⚡ Google launches Gemini 4 Neon
🦠 Microsoft's Python AI framework hijacked to steal credentials
🔓 UK finds GPT-6.2 Astra easier to jailbreak
📄 Cloudflare open-sources two data extraction models
Plus: 💡 5 strategies & tactics, 🎁 8 more stories you might like, 🧰 6 tools, and 📚 5 papers.
Google launches Gemini 4 Neon LINK
- Google launched Gemini 4 Neon, a mid-range model, earlier today on the paid Gemini API in AI Studio, Android Studio, the Gemini app, and Antigravity for Pro and Ultra subscribers.
- At $3 per million input and $15 per million output tokens, Neon beats Opus 5.5 within about a point on two engineering benchmarks while running roughly 1.9x faster than Carbon at ~110 tokens/second.
- Its million-token window ingests video, audio, charts, spreadsheets and PDFs together, but at standard rates Neon must use about one-third fewer tokens than Sol or Sonnet to erase its 50% price premium over them.
Microsoft's Python AI framework hijacked to steal credentials LINK
- Microsoft's official agent-framework package on PyPI was hijacked to install info-stealing malware, security researchers reported yesterday and today, after an attacker appears to have gained access to an administrator account on Microsoft's GitHub organization.
- The two malicious versions of the Python package, which sees nearly 200,000 downloads a month, exfiltrate secrets for Azure, Git, Docker, and Hugging Face without ever being imported, just an installed package and a running interpreter trigger it.
- Beyond passive theft, the malware actively requests fresh identity tokens from GitHub Actions, Azure DevOps, and Azure, sending them to a hardcoded webhook, so anyone finding either version should assume account cleanup is needed, though the code isn't persistent and runs only while the malicious package remains installed.
UK finds GPT-6.2 Astra easier to jailbreak LINK
- The UK AI Security Institute built a universal jailbreak against two pre-release versions of GPT-6.2 Astra using half the model queries needed to break GPT-6 Astra, OpenAI disclosed in a Nov. 17 system card.
- Red-team lead Xander Davies said a single researcher developed the cyber jailbreak against both versions and extended it to agentic tasks, eliciting substantial capabilities though preliminary tests found no edge over leading open-source models.
- OpenAI revised the model's safety instructions and gave the institute three extra business days to retest, saying internal testing mitigated the regression, though the institute had not independently verified the changes when the card was written.
Cloudflare open-sources two data extraction models LINK
- Cloudflare released open weights for two data-extraction models on Hugging Face, the 27B Clairic and a new 9B Clairic-flash, both turning unstructured text, images and PDFs into typed values, estimates and forecasts, also hosted on Workers AI.
- Built on Qwen3.8-27B and Qwen3.5-9B under Apache-2.0, they swap the preview's single-forward-pass for parallel slot autoregressive generation with separate heads for estimates, forecasts and arithmetic, improving text and image accuracy at a small latency cost.
- Standard Workers AI pricing is $0.24 per million input tokens for Clairic and less than half that for Clairic-flash with no output charge, both discounted until December 1, 2026, though Cloudflare calls its accuracy comparisons directional and self-reported, with no public extraction benchmarks to rank against.
Strategies & Tactics
> Design Once, Deploy at Scale: Building a Template-Driven ML Ecosystem: Standardizing machine learning models into a few reusable templates lets Meta improve one template and automatically propagate that gain across thousands of ads-ranking models, cutting per-model engineering time by 92%.
> Your AI coding agent could cost 5x more depending on the harness, benchmark finds: The scaffolding wrapped around an AI coding model can swing running costs fivefold, so testing harness-model pairings matters as much as picking the model.
> Why your fine-tuning job takes longer than the FLOPs say: A machine-learning model trained on past runs predicts fine-tuning job time far more accurately than FLOP math, because it captures overheads like evaluation and checkpointing that compute-only estimates miss.
> 200 OK is dead as an AI agent health signal. Here’s what comes next.: Monitor AI support agents by adding a second classifier that catches polite refusals wrongly turning away valid customers, failures that standard health checks mark as success.
> Evals for AI Engineers: Diagnose AI agent failures by pinpointing which of three gaps, misreading the data, vague instructions, or inconsistent model behavior, is the real cause before fixing.
Other news you might like
- Anthropic adds Code Mode to Claude Code for chained JavaScript tool calls LINK
- Reson8 releases Rumble-2-Flash speech model as a free open-weights download LINK
- Microsoft-Decision-1.1 is now generally available for fast, low-cost decision making LINK
- Snapchat now lets developers build AR Lenses with AI assistants like Claude Code and Cursor LINK
- AWS Brings Full-Weight Training To Serverless SageMaker Customization LINK
- US lawmakers seek Microsoft briefing on reported AI deal talks with ByteDance LINK
- Horizon Labs Report LINK
- Introducing eve 1.0 LINK
Trending tools
TestSprite 4.0: an AI agent that automates frontend and backend testing from planning to execution, using natural language to cut testing time and costs. LINK
Replay: records Mac meetings locally across Meet, Zoom, Teams, and Slack huddles, producing searchable transcripts and notes with no bot joining, for a one-time $10 LINK
Teaflask: adds a managed AI copilot to your existing SaaS via the frontend, letting users complete tasks by asking without rebuilding your backend. LINK
Spellar for Chrome: records meetings without a bot, generating summaries and action items, chatting across past calls and sending tasks to Notion, Linear, and Todoist. LINK
Tecla: a proactive AI assistant that scans your email, Slack, and meetings to auto-build a to-do list, then completes tasks when you click Run. LINK
Gemini 4 Hydrogen: multimodal AI model handling text, images, audio, video, and code, built for fast coding and agentic workflows across Google's developer tools. LINK
Trending papers & reports
Introspective oversight gives a model a built-in watchdog role that answers honest questions about the assistant's hidden misbehavior, catching deception and reward-gaming more accurately than external monitors across five tests. LINK
AI agent safeguards stop interrupted tool actions from being duplicated or wrongly assumed done, cutting repeated effects to zero from 117 of 200 cases, though the safety layers run about seven times slower. LINK
Higher coding benchmark scores did not translate into more real work, as a top model boosted developers' merged code by just ~1.4 points, a result too small to be statistically meaningful. LINK
AI cost-cutting setup uses an automated agent to custom-design each model's compression recipe in about an hour, beating the best public settings by up to ~2 points of accuracy while running on existing hardware. LINK
Clover language model tops seven coding agent benchmarks while generating 680 words per second on a single chip, and answers batches of related questions together to lift accuracy from ~90% to ~94%. LINK
See you tomorrow for a new dose of ☕️ AIpresso!