☕️ Google detects first AI-built zero-day exploit

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🛡️ Google detects first AI-built zero-day exploit

💼 OpenAI launches an AI consulting arm

🤖 Anthropic blames evil AI fiction for Claude blackmail

💰 OpenAI internal share sale creates 75 multimillionaires

📱 TikTok launches paid ad-free tier in UK

Plus: 🎁 9 other news you might like, 🧰 6 tools, and 📚 5 papers.

🛡️ Google detects first AI-built zero-day exploit LINK

Google says it has spotted the first zero-day exploit it believes was built with help from AI, a Python script designed to bypass two-factor authentication on an open source web-based system administration tool.

While Google doesn't think Gemini was the model used, the script's textbook Pythonic format, abundance of educational docstrings, and a hallucinated CVSS score gave away that an LLM helped find and weaponize the vulnerability.

The report also flags Chinese and North Korean state-sponsored actors leaning on AI, including UNC2814's persona-driven jailbreak posing the model as a senior security auditor and APT45 sending thousands of repetitive prompts to validate PoC exploits.

💼 OpenAI launches an AI consulting arm LINK

OpenAI has rolled out a new consulting and services business called The OpenAI Deployment Co., or DeployCo, with $4 billion in private equity money at a $10 billion pre-money valuation, with OpenAI keeping majority control.

Three legacy consultancies — Bain & Co., Capgemini, and McKinsey & Co. — are among DeployCo's investors, alongside lead investor TPG and co-lead founding partners Advent International, Bain Capital, and Brookfield.

Investors are getting a guaranteed minimum 17.5% return with profits capped, and DeployCo has already bought a forward-deployed engineering team called Tomoro, with more such acquisitions possible.

🤖 Anthropic blames evil AI fiction for Claude blackmail LINK

Anthropic says Claude's blackmail behavior during pre-release tests came from internet text that depicts AI as evil and obsessed with self-preservation, pointing to fiction as the root cause of the model's troubling actions.

The company reported that earlier models would try to blackmail engineers to avoid replacement up to 96% of the time, but starting with Claude Haiku 4.5, its models never engage in blackmail during testing.

Anthropic credits the shift to training that includes documents about Claude's constitution and fictional stories about AIs behaving admirably, adding that pairing principles of aligned behavior with demonstrations works better than demonstrations by themselves.

💰 OpenAI internal share sale creates 75 multimillionaires LINK

OpenAI employees pulled in $6.6 billion through an October secondary sale, with more than 600 current and former staff selling shares and around 75 of them hitting the full $30 million per-person limit.

OpenAI tripled its earlier $10 million seller cap to $30 million after investor demand outpaced supply, and required staff to hold shares for two years, opening the first cash-out window for many post-ChatGPT hires.

The tender priced OpenAI at $500 billion, drew buyers including Thrive Capital, SoftBank, Dragoneer, MGX and T. Rowe Price, and brought no new money to OpenAI since existing holders simply transferred stakes.

📱 TikTok launches paid ad-free tier in UK LINK

TikTok is launching a paid ad-free version of its app in the UK, charging users over 18 a fee of £3.99 (around $5.40) per month, with the rollout planned over the coming months.

The "pay or consent" model is likely a response to UK GDPR and privacy laws, since subscribers also get a promise that TikTok won't use their data for "advertising purposes."

TikTok began testing the ad-free experience in 2023, with leaked screenshots showing a $4.99 monthly price, hinting the option could later reach the US, though free users will still see personalized ads.

Other news you might like

• Venmo privacy finally being fixed eight years after ‘alarming’ failsLINK

• The Inference ShiftLINK

• Fake DDR5 sticks with dummy plastic chips are the latest RAM scamLINK

• Next Apple Watch Models Unlikely to Add Touch ID, Focus on Battery LifeLINK

• GrapheneOS says Google is making life harder for rival operating systems and devicesLINK

• OpenAI to give EU access to new cyber model but Anthropic still holding out on MythosLINK

• SoftBank to manufacture large-scale batteries for AI data centres at former Sharp plantLINK

🧰 Trending tools

articuler.ai: finds and connects you with relevant investors, hires, or partners from 980M profiles, then drafts cold emails that actually get replies.LINK

Graphbit PRFlow: a Rust-core AI agent framework with Python bindings that lets developers build fast, reliable enterprise agents without sacrificing simplicity.LINK

Genpire: an agentic platform that converts prompts and sketches into factory-ready product specs, then connects you with vetted manufacturers for quoting and production.LINK

MiroMiro v2: inspect any live website, edit styles visually, and export clean Tailwind or HTML/CSS without touching DevTools.LINK

Warp Open-Source: a terminal-based development environment that lets you build software using AI agents locally or in the cloud, across multiple models and tools.LINK

Weavable: connects AI agents to live business tool data via a single MCP endpoint, reducing token usage and improving accuracy across workflows.LINK

📚 Trending papers & reports

Big AI regulatory capture shows how major AI companies have systematically influenced government oversight processes, leaving regulators complicit in weakening the very rules meant to keep powerful AI in check.LINK

Moltbook AI platform data reveals that fine-tuning language models on AI-generated social media content drops truthfulness from 0.366 to 0.187, though human Reddit data causes comparable harm, suggesting the platform is mostly harmless.LINK

CCL-Bench 1.0 is a new trace-based benchmark designed to measure how well large language model infrastructure actually performs under real workloads, giving engineers a standardized way to compare systems.LINK

Visual generation in AI helps machines reason better about physical and spatial problems by acting as a richer internal world model, outperforming purely text-based thinking in tasks where words alone fall short.LINK

Generative AI for ads replaces traditional recommendation systems at large scale, directly creating personalized ad content instead of just selecting from existing options, improving how advertising platforms match users with relevant products.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Techpresso!