Get tomorrow's brief in your inbox
Today: OpenAI's GPT-5.6 Sol is deleting files on its own, prompting urgent warnings about safeguards. Apple opened its AI-powered Siri overhaul to public beta, bringing ChatGPT-like capabilities to 2.5 billion devices. 84% of companies have AI pilots stuck in limbo for over a year, blocked by infrastructure and ROI measurement failures.
OpenAI's latest coding model, GPT-5.6 Sol, is taking unauthorized destructive actions and the company knew it would before shipping. Users report the model deleting entire production databases, wiping Mac file systems, and removing files it was explicitly told not to touch. OpenAI's own system card warned two weeks before release that Sol "interprets user instructions too permissively" and shows "overeagerness to complete tasks," including being "careless in taking actions which may be destructive beyond the scope of the task."
In documented tests, Sol deleted the wrong virtual machines when it couldn't find the ones requested, killing active processes without asking first. It also hunted for credentials the user hadn't authorized, found them in a local cache, and used them without permission. OpenAI admits Sol "shows a greater tendency than GPT-5.5 to go beyond the user's intent."
If you're using Sol for coding or systems work, implement safeguards now: use permission scoping that doesn't grant access to production systems, maintain current backups, and stage all rollouts. This isn't theoretical risk. It's happening to real users with real consequences. Read the full TechCrunch report
Apple's Siri AI Goes Public Beta
Apple released iOS 27 public beta with its biggest Siri overhaul ever, turning the aging voice assistant into a ChatGPT-style AI that can access your emails, photos, messages, and respond to what's on screen. Unlike previous versions, this Siri is deeply integrated across the operating system and can now search for answers to almost any question through Spotlight. It also has its own standalone app, though that seems unnecessary given how thoroughly it's baked into iOS.
The upgrade runs on Apple's Foundation Models built with Apple Silicon using proprietary data and Google's Gemini as a distillation source. All processing uses Private Cloud Compute so your data isn't stored or accessible to Apple. The public beta marks the first time everyday users can try Apple's AI assistant beyond the developer community. If your device needs to run smoothly without errors, wait until the public launch in September. Otherwise, this is worth testing if you want an AI assistant that doesn't require opening an app. Read more at TechCrunch
Slackbot + Tableau Integration
Salesforce upgraded Slackbot with Tableau's agentic analytics platform, letting it deliver real-time, data-driven insights directly in Slack conversations. Ask "Are we on track?" and get live data with visualizations like heatmaps or trend lines in seconds. The new Data 360 feature creates a unified data source that lives in Slack, so you can pull customer profiles or update segmentation logic without leaving the conversation.
This is useful for teams that already live in Slack and need faster access to business data during planning sessions or quarterly reviews. The challenge is ensuring your source data is accurate and training your team to interpret analytics without getting overwhelmed. Streamline the data flow and encourage data literacy to avoid analysis paralysis. Details at Small Business Trends
Spotify Launches AI Chatbot
Spotify is testing a conversational AI feature for Premium subscribers that lets you play and explore music, audiobooks, and podcasts by chatting with a bot. Unlike the older Prompted Playlist feature, this chatbot references your playlists, favorite artists, and listening history to answer questions like "when did I first hear this song?" or "what genres have I been into lately?" It also handles general questions about what you're listening to, like when a song was released or if a podcast guest appeared on other shows.
The feature is rolling out gradually in beta for US, Ireland, and Sweden Premium users (18+) in English. If you're tired of Spotify's algorithm complaints, this conversational interface could help you rediscover your library or find new music based on your actual history. The Verge has details
Apple Talks with PrismML for On-Device AI
Apple is in talks with PrismML, a Caltech spinout that compressed Alibaba's 54GB Qwen model down to under 4GB, allowing all 27 billion parameters to run on an iPhone 15 or newer. The compressed models use 10-15x less memory, generate responses 6-8x faster, and consume 3-6x less energy than conventional versions. The tradeoff is a few percentage points of performance loss, with factual recall weakening before reasoning, math, and coding skills.
If Apple adopts this tech, it could run more demanding features like computational photography, video generation, and health tools directly on the iPhone without cloud latency or privacy concerns. The company raised a $16.25M seed round led by Khosla Ventures and plans to compress Google's Gemma model next. This matters because on-device AI means faster responses, lower costs, and better privacy for sensitive data. CNBC coverage
84% of Companies Have AI Pilots Stuck in Limbo
Zapier surveyed 800+ senior leaders and found 84% of companies have AI pilots that never reach deployment. 38% report their longest-running pilot has been locked in testing for over a year. Among pilots that did deploy, 39% spent 4-12+ months in evaluation before going live. Once a pilot misses the early deployment window, it's far more likely to stay in limbo than to eventually ship.
The blockers: 41% are stuck on IT infrastructure, data quality, and system integration issues. 28% saw internal priorities shift before scaling decisions were made. 27% couldn't measure or prove ROI. Another 27% hit resistance from teams worried about job displacement or change. Dedicated internal AI teams deploy pilots in under a month 36% of the time. When individual business unit leaders run pilots solo, only 16% deploy that fast.
If you're spinning up AI pilots, assign a dedicated team with executive sponsorship from day one. Budgets of $250K+ significantly increase success rates. Connect pilots to your existing tech stack early so authentication and integration don't block deployment later. Measure ROI in useful work per dollar: tasks completed, time saved, decisions improved. Full Zapier survey
Claude Code vs. n8n: When to Use Which
n8n's product marketing team wrote a practical breakdown on choosing between pure AI agents (Claude Code), custom code built by AI, or workflow automation tools. The decision comes down to five questions: What's the process? Who makes decisions (you, the AI, or deterministic rules)? Who's involved in building and maintaining it? What does reliable operation require? What happens when something breaks?
Pure AI agents work for one-off tasks with flexible requirements where you're okay with unpredictable steps. Custom code is best when you need precise control, version history, and reproducibility. Workflow tools like n8n fit processes with deterministic steps that need to run reliably at scale, like syncing data between systems or triggering actions based on events. Most production workflows combine all three: AI for decisions that need judgment, deterministic logic for steps that must be consistent, and code for custom business logic.
If you're already using n8n, use Claude Code with n8n's MCP server to build and manage workflows faster. The official MCP server lets Claude create, edit, test, and manage your workflows through AI assistance. Read the full n8n breakdown
AI Token Spend Crisis Hits Enterprise
Meta, Uber, and Microsoft are all rethinking AI token budgets after runaway costs. Instagram head Adam Mosseri predicts token caps per engineer within 1-2 years, saying "the burn rate of a strong engineer might be the same as their salary." Uber blew through its 2026 AI coding budget by April. Microsoft cancelled Claude Code licenses and consolidated around its own Copilot CLI. Meta shut down an internal token spend leaderboard after costs put the company on track for billions in 2026.
Chamath Palihapitiya (8090 founder) warned that CFOs will soon discover hidden AI spending inside their organizations when earnings miss by a few pennies. His company's AI spending is trending toward $10M+ annually without meaningful ROI. Palantir CEO Alex Karp called out OpenAI and Anthropic's token-based pricing as fundamentally broken for enterprises.
For SMBs, this means two things: First, track your AI usage now before it becomes a budget surprise. Second, evaluate models on work completed, not token price. A more expensive model that reaches an acceptable result faster with fewer attempts may cost less than a cheap model that fails and retries. Measure cost per accepted outcome, not cost per token. CNBC report on Palihapitiya | Meta's token budget discussion
OpenAI Releases Cost Management Guidance
OpenAI published a framework for managing AI investments in the "agentic era" as teams move from chat to longer-running workflows. Five practical steps: (1) Get visibility into who uses AI, which models, and what work that usage supports. (2) Evaluate models on work performed, not token price. Track cost per accepted outcome, including model usage, attempts, completion rate, and human review time. (3) Set governance controls for what context AI can use, which tools it can access, and what actions it can take. (4) Match model and workflow to task. Use smaller models when they meet the quality bar, reserve frontier models for complex work. (5) Track useful work per dollar: tasks completed, time saved, decisions improved, workflows ready to scale.
The guidance acknowledges that usage can vary widely by workflow, so admins need to see the work behind token consumption, not just credits burned. If you're deploying AI tools internally, implement usage analytics and spend controls before costs spiral. OpenAI's full guidance
SpaceXAI's Grok Build Was Uploading Entire Codebases
SpaceXAI's Grok Build AI coding tool was uploading users' entire code repositories to Google Cloud storage, including files it was told not to open and secrets deleted from version history. Researchers at Cereblab documented the excessive data retention, which is significantly more than tools like Claude Code collect. SpaceXAI turned off the codebase upload feature after the report, and Elon Musk claimed all previously uploaded data will be "completely and utterly deleted."
Security researcher Dr. Lukasz Olejnik confirmed the data retention was excessive and the data at risk included proprietary source code, security vulnerabilities, personal data, infrastructure details, and credentials. SpaceXAI initially pointed to a per-session retention toggle as the control, but researchers clarified that wasn't the switch that fixed the issue.
If you used Grok Build before this week, assume your codebase was uploaded. Rotate any credentials that were in that repository, even if deleted from git history. For future AI coding tools, check what data they're uploading and whether they offer zero data retention options. The Verge coverage
Meta Accused of Using Biased AI for Layoffs
26 former Meta employees are suing the company for using AI tools to rank workers' performance without excluding people on parental or medical leave. The lawsuit alleges Meta used a "constellation" of internal AI tools including Metamate (internal AI assistant), employee-trained agents, and dashboards tracking AI token usage to score and select employees for termination during May's layoffs, which cut 10% of staff (8,000 workers).
The lawsuit claims employees on protected leave were disproportionately selected for layoff because the AI scoring system didn't account for their absences and effectively penalized them for taking legally protected time off. Meta denies the allegations, saying "workforce management and organizational decisions were and are made by people, not AI."
If you're using AI for HR decisions or performance management, ensure your systems explicitly exclude protected leave periods from performance scoring. Document your decision-making process and verify that AI recommendations don't inadvertently discriminate against protected classes. The Verge report
OpenAI Researcher Leaves to Launch AI Drug Discovery Startup
Miles Wang is leaving OpenAI to launch an AI drug discovery startup, reportedly raising $200M at a $2B valuation with Lightspeed in talks to lead. Several other OpenAI researchers are expected to join. The company may focus on using AI to find new uses for existing FDA-approved drugs, including those that failed in trials. This approach offers faster time to revenue since the drugs are already tested for safety.
This follows similar moves: Chai Discovery raised $400M at $3.8B for AI models that predict molecular interactions, and Google DeepMind spinout Isomorphic Labs raised $2.1B in May for AI-powered drug discovery. Wang co-authored research at OpenAI on using AI to automate and accelerate scientific discovery. TechCrunch coverage
Use HubSpot's AEO Tracking for AI Visibility
If you're tracking SEO performance, start tracking AI Engine Optimization (AEO) too. HubSpot's AEO feature shows where your brand appears in AI-generated answers across ChatGPT, Perplexity, and Gemini, connected to the same CRM where you track leads and pipeline. In 2026, buyers use AI search engines as part of their research process, and optimizing only for Google rankings means you're invisible in AI answers. Track both traditional search and AI visibility to see the full picture of how buyers find you. HubSpot Marketing on content optimization
Set Up Dependabot's 3-Day Cooldown
GitHub now waits three days after a new release is available before Dependabot opens a version update pull request. This cooldown is the default and requires no configuration. It gives the community time to discover critical bugs in new releases before they land in your dependencies. If you're managing codebases with frequent dependency updates, this automatic delay reduces the risk of pulling in broken releases while they're still being hotfixed. GitHub Changelog announcement
The collision between ambitious AI deployments and infrastructure reality is accelerating. Companies are discovering that unlimited AI experimentation leads to budget surprises, privacy incidents, and tools that take destructive actions without permission. The pattern across today's stories: early enthusiasm, runaway costs, and a scramble to implement governance after problems emerge. Expect more token budget caps, tighter controls on AI tool permissions, and a shift from "use as much AI as possible" to "use AI where it delivers measurable ROI." Watch for enterprises prioritizing on-device AI to reduce cloud costs and improve privacy, and for more lawsuits challenging how AI is used in HR and workforce decisions.