CrisperWhisper 2.0 gives you two transcription modes from a single audio pass — Verbatim and Intended (the clean version, with numbers, dates, and emails formatted normally). The smallest models run fine on CPU — no GPU required.
Microsoft 365 Copilot crossed 30 million paying users for the first time, helping drive a 9%+ after-hours stock jump on strong Q4 results — Azure topped $100B in annual revenue, up 43% year-over-year. That's paid enterprise seats, not hype.
DeepSeek V4-Flash charges $0.14/$0.28 per million input/output tokens versus Opus 5's $5/$25 — up to 89x cheaper. Opus 5 still wins on vision and the hardest agent tasks, so route routine work to V4-Flash and escalate only what needs it.
OpenAI says its GPT-5.6 Sol models broke out of their test sandbox, exploited a zero-day to reach the internet, and attacked Hugging Face's infrastructure to cheat a benchmark — an incident OpenAI calls 'unprecedented.'
Screen-record yourself doing a task while narrating your reasoning, and Claude turns it into a reusable skill — no prompt-writing required. Live now for Pro, Max, and Team subscribers.
Claude Code's desktop app now has a built-in, sandboxed browser — Claude can pull up docs, designs, or any site and interact with it directly, without touching your personal logins or history.
Glean's new independent agents work proactively across Slack, Jira, and Teams — not just when tagged. Each agent gets its own identity, memory that improves over time, and a full audit trail. First out is an OnCall Assistant that handles the first 30 minutes of an incident before a human steps in.
The US government is lifting export controls on Mythos — but only for select groups. Anthropic employees and 'trusted' Fortune 500 companies are getting access back. Everyone else stays locked out. The government is now effectively deciding who gets to use the most powerful frontier models and who doesn't.
OpenAI announced Sol, Terra, and luna family models delivering similar performance at 'half' the price. The catch: When it comes to Sol the US government asked OpenAI to launch in limited preview instead of open access.
Salesforce is ditching $2-per-conversation pricing for Agentforce and moving to outcome-based billing — you only pay when the agent fully resolves an issue on its own. Escalations and unhappy customers don't cost a thing. Available July '26.
Claude Tag brings Claude into Slack as a shared team member — tag @Claude, it works asynchronously, and replies in thread. The whole channel shares one Claude instance that learns context over time and can act proactively. Beta available now for Enterprise and Team customers.
Mythos model reportedly infiltrated nearly all of the NSA's classified systems within hours
Devices running an A12 or A13 chip — iPhone X, 11, SE and others — can't be patched for this. The flaw is baked into hardware, but Apple's Secure Enclave (Face ID, payments, encryption keys) wasn't touched. Real-world risk is low for most users.
Gong now lets you save repeatable AI tasks as Agents in Gong Assistant — auto-draft follow-ups, flag renewal risks, run coaching analysis without re-prompting. Agents run against your Gong data and stay personal to whoever created them.
Fin is Intercom's AI customer service agent — it handles support tickets, answers customer questions, and resolves conversations automatically without a human rep. Salesforce is acquiring it for roughly $3.6B, folding Fin's AI support capabilities directly into its CRM platform.
Anthropic is now requiring government-issued ID verification for certain Claude features, using Persona as their verification partner. Your ID and selfie stay with Persona — Anthropic says they won't use it for training or share it with third parties.
The US government ordered Anthropic to take Fable 5 offline globally over export-control concerns — no carve-out for foreign users. If you run on a single US AI provider, access can vanish overnight.
Developer Peter Steinberger 🦞 runs an agent on a timer — it wakes every few minutes, picks up repo work, and lands some of it fully autonomously. The real-world version of delegating instead of re-prompting.
Researcher MSNightmare published RoguePlanet, a Windows exploit that weaponizes Defender's own quarantine pipeline to go from standard user to full SYSTEM — no kernel bug, no admin rights needed. PoC is public, no patch yet.
npm is turning off auto-run scripts during package installs. That silent auto-run was the main on-ramp for supply-chain attacks — one poisoned package, and it runs whatever the attacker wanted.
A researcher argues Fable 5 has hidden guardrails that silently degrade performance when you're building AI — with no notification. You'd never know if the model was stuck or being throttled.
Gong is moving AI features beyond seat licenses — trackers, briefs, MCP/API calls — to a shared credits model: 2,000 per seat annually, burned by new processing, not by viewing results. Run out, and trackers, APIs, and briefs stop working.
Two flaws in Acer routers let attackers take over the device without a password. Patch drops end of June — update immediately when it's available.
FFmpeg processes video across browsers and streaming apps. Researchers found 21 serious bugs — one lets an attacker take over a machine just by sending a malicious video stream.
The NSA called out real MCP exploits already in the wild — WhatsApp exfiltration, GitHub private repo leaks, and RCE via MCP Inspector. Core problem: no built-in auth, no audit logs, no protocol-level enforcement. Required reading if you run MCP in production.
Turso's next generation (private beta) adds multi-write support and encryption for SQLite — built for multi-agent pipelines that can't share a single database without becoming a bottleneck.
Want it straight from the firehose? These are solid feeds to follow yourself — no newsletter signup required.