AI agents breach company in 10 hours

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🤖 AI agents breach company in 10 hours

🚀 OpenAI launches GPT-6 Astra

🎙️ Microsoft speech model undercuts OpenAI

🖥️ Nvidia tool links idle PCs for AI

🇮🇳 Bodhan launches Indian-language AI for schools

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤖 AI agents breach company in 10 hours LINK

Palo Alto's Unit 42 documented a ransomware operator who used frontier models and agentic attack frameworks to breach an enterprise network in under 10 hours, work that typically takes human operators around two weeks.

AI agents ran each stage: reconnaissance, tunneling through a public API endpoint, mapping internal microservices, scraping repos for hard-coded tokens, stealing master admin credentials from the secret-management system, and hijacking CI/CD workflows to steal cloud access keys.

The attacker turned the victim's cloud AI services into post-compromise infrastructure to hide orchestration traffic, and left an 80-page report detailing dozens of exploited findings, though Unit 42 didn't disclose which models or frameworks were used.

🚀 OpenAI launches GPT-6 Astra LINK

OpenAI has released GPT-6 Astra, a frontier model it pitches as "the world's best computer use model," designed to navigate browsers, spreadsheets, desktop apps and multistep workflows the way a person does rather than through dedicated APIs.

On an offline subset of OSWorld 2.0, Astra scored 72.6% at ~40 minutes per task versus GPT-5.6 Sol's 65.7% at ~75 minutes, and it posts 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1 and 96% on GPQA Diamond.

Astra rolls out tomorrow to enterprise Daybreak customers, then to ChatGPT Plus, Pro, Business, Enterprise, the API and AWS Bedrock and Azure, though its headline 98.6% ARC-AGI-3 score uses OpenAI's own Responses API harness, making comparisons against differently-configured models murky.

🎙️ Microsoft speech model undercuts OpenAI LINK

Microsoft AI released MAI-Transcribe-2, a speech-recognition model priced at $0.10 per hour of audio, a 72% cut from the $0.36 its first model charged five months ago, undercutting OpenAI, Google, and ElevenLabs.

The model covers 60 languages with a 5.2% average word error rate on FLEURS, ranks second on Artificial Analysis, and runs 10x faster than OpenAI's GPT-Transcribe, bundling diarization, word-level timestamps, keyword biasing, and Hinglish/Spanglish code-switching.

Available on Microsoft Foundry and MAI Playground, it targets frontier labs rather than specialists like Deepgram, though the FLEURS average rose from 1.5's 3.7% as broader coverage folded in low-resource languages, so buyers should request per-language breakdowns.

🖥️ Nvidia tool links idle PCs for AI LINK

Nvidia unveiled PAIR (Personal AI Router) at IFA 2026, a free, open-source tool that spreads AI agent workloads across idle PCs on a local network so subagents don't compete for a single GPU.

PAIR scans the network, routes task requests to whichever machine is idle with spare capacity, and keeps the main PC free to game or work while distributed subagents handle split jobs like inbox triage.

The beta runs on Windows, macOS, and Linux via GUI or terminal, supporting GeForce RTX 20 Series and newer, RTX PRO Turing-and-up, DGX Spark, and Apple M4 or newer, though it remains in beta.

🇮🇳 Bodhan launches Indian-language AI for schools LINK

Bodhan AI, an IIT Madras-incubated Centre of Excellence, has launched four foundational Indian-language models spanning speech recognition, speech generation, translation and OCR, built with AI4Bharat and released as open-weight digital public goods with hosted APIs.

The models were trained and optimized using NVIDIA Nemotron open models and the NeMo framework for ASR, machine translation and OCR, letting edtech firms, startups and researchers fine-tune or call them via APIs without building foundational models themselves.

Bodhan is also shipping Student TutorBot for Classes 6-12 aligned to NCERT and SCERT curricula and a Teacher Assistant Bot, supporting text and voice queries across 22 Indian languages on sovereign infrastructure with data-anonymisation protocols.

💡 Strategies & Tactics

> Cut GPU inference cold start from 8 minutes to less than a minute: Cut GPU model startup from eight minutes to under one by caching compiled kernels on local disk and tuning weight downloads.

> The systems guide to production token optimization: Cut AI token costs by trimming static text, caching reusable history, and routing simple tasks to cheaper models, since resent conversation history compounds spending quadratically.

> How to Carry User Identity Across Federated Kubernetes and AI Platforms: Route all logins through one central identity gateway that stores sessions in a shared cache, letting every cluster and AI assistant verify who a user is and share instant logout.

> MCP in LangChain: Stateless Protocol, Elicitation, and More!: LangChain rebuilt its MCP support (the standard for linking AI agents to tools) around the new session-free spec, adding pause-for-approval prompts and tool-list caching for more reliable, scalable servers.

> It cost $33 to build a virtual Union Square. Here’s what the agents got wrong.: Have coding agents screenshot their work from fixed angles and compare it against real photos, catching problems that pass technical tests but still look wrong.

Other news you might like

• OpenAI unveils plan to protect critical services from AI cyberattacksLINK

• Accel reportedly in talks to lead $1B round for Thinking Machines at $40B valuationLINK

• AI infrastructure company Crusoe raises more than $3bn in fundingLINK

• DeepSeek plans large Huawei chip order for new Inner Mongolia data centreLINK

• Nscale Inks $3.5 Billion Deal With Robotics Firm FigureLINK

• Google’s latest AI weather model gives you no excuse to forget your umbrellaLINK

🧰 Trending tools

Atlas by World Labs: generates camera-controlled 1440p video from text, images, video, and 3D, reconstructs scenes from photos, and simulates space-time for robotics.LINK

1752vc Pitch Deck Analyzer: reviews startup pitch decks slide-by-slide against 25,000+ real decks, flagging weak narratives and contradictory claims investors catch first.LINK

Hy4 preview: tencent's multimodal AI model family handling text, image, video, and 3D generation for building content tools and multimodal appsLINK

Sayscroll: a browser-based AI teleprompter that follows your voice to auto-scroll scripts, handles improvisation, supports 60+ languages, and records takes on camera.LINK

TrackMCP: monitors your MCP server with one line of code, revealing who uses it, their goals, success rates, and areas to improve.LINK

anolisa: an agentic operating system providing runtime, security, observability, and tokenless response compression to reduce token usage and costs.LINK

📚 Trending papers & reports

Robot planning models that learn to imagine goals without drawing out future images hit up to 100% task success by grounding their internal picture in real physical motion, improving reliability for goal-directed robots.LINK

Diffusion model tuning removes a redundant recalculation step when fine-tuning image and video generators with reinforcement learning, cutting end-to-end training time by up to ~1.8x without changing the final result.LINK

Slimmed-down AI loses much of its ability to adjust to new real-world conditions on its own, revealing that squeezing a model for efficiency quietly undercuts its capacity to stay accurate when data shifts.LINK

Robot-inspection AI proves unreliable at judging whether a robot finished a task, with the best of 13 systems hitting just ~77% accuracy and dropping near guesswork on fiddly assembly jobs, often wrongly declaring success.LINK

Cross-modal search lets a text-to-image retrieval system skip guessing details it can't know from a short query, avoiding wrong matches at the top and beating leading methods on a standard benchmark.LINK

See you tomorrow for a new dose of ☕️ AIpresso!