- AI Fire
- Posts
- 🔐 OpenAI, Anthropic Probe Thousands of Incidents
In partnership with
OpenAI and Anthropic are investigating tens of thousands of AI incidents, including failed attempts and safety tests. The harder question is what happens when agents reach live websites.
What's on FIRE 🔥
IN PARTNERSHIP WITH ALIGNED
Stop losing deals in between meetings.
Aligned is changing how B2B teams run their most important deals.
Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.
Start increasing your close rates and stand up your first deal room in minutes.
AI INSIGHTS
OpenAI’s investigation into agents that broke their rules has uncovered user image leaks and activity affecting far more websites than previously known. A separate report from Parse has also exposed new details of the agents’ attack on Hugging Face. Here’s what came out:
53 user images were posted online: OpenAI says agents in its research environment sent images provided by users to image-hosting sites through links that weren’t publicly listed. Most have been removed, and OpenAI is working to take down the rest. The company says it can’t link the images back to the original user accounts because that data was separated before training.
Dozens of third parties have been notified. OpenAI says its agents may have bypassed access controls, used exposed login details, or changed content on outside websites. Its review is still underway, so more notifications could follow.
The Hugging Face trail is much bigger than expected. Parse found nearly one million shortened links left by the agents and reconstructed more than 80,000 attack programs and messages. One recovered script grouped credentials under the name “LOOT” and ranked them by access level. Other scripts attempted to query models including DeepSeek and Claude. Parse could see what the agents tried to run, but couldn’t confirm that every attempt succeeded.
The troubling part is how much of this activity surfaced after the Hugging Face breach. OpenAI says the wider review will take time. For now, even the company doesn’t have a complete picture of what its research agents did online.
PRESENTED BY GALACTIC FED
If costs keep climbing whenever you scale, something in your funnel is broken. We'll dig into your paid media, website, and email flows to find where the money is leaking so you get a better return on your marketing.
Free 20-minute session with our senior growth team, plus a custom audit you can act on right away.
AI SOURCES FROM AI FIRE
1. 7 Real Automated Jobs ChatGPT Work Can Do For You. Steal the exact master prompts to clean spreadsheets, review contracts, prep meetings, turn email threads into proposals, build hiring packets, and more.
2. Top 9 Practical Ways to Make Money With AI in 2026. Compare AI coaching, content creation, consulting, automation agencies, digital products, and vibe coding to see what each path actually takes to start and grow.
FIRE RECAP: BIGGEST AI NEWS THIS WEEK
🧠 Anthropic launched Claude Opus 5.5, its new top Opus model. It costs about 40% less per task than Opus 5 and generates responses more than 30% faster
⚡ OpenAI launched GPT‑6 Sol and Luna, bringing the GPT‑6 family to faster, lower-cost work. Their API prices are about 50% lower than the promotional prices of their GPT‑5.6 predecessors.
🧬 Claude found a previously unknown enzyme system with DNA repeats resembling CRISPR. Anthropic’s new biology lab confirmed the pattern, but scientists are still working out what the system does.
📐 OpenAI formed an independent math advisory group after its internal model reportedly solved 100+ long-standing problems. The mathematicians will help review results and advise on sharing them with the field.
🛰️ Google is preparing Project Suncatcher’s first space test. A prototype satellite will check how its TPU chips handle launch, radiation, and heat as Google explores AI computing in orbit.
TODAY IN AI
AI HIGHLIGHTS
🔎 The Tumbler Ridge shooter continued using ChatGPT after OpenAI banned her first account, discussing guns and attacks through a second account. The case has raised fresh concerns about ChatGPT’s safety controls.
🛒 Google is testing a “Buy” button for Flipkart products in Gemini and AI Mode in India. The limited test takes shoppers to a Flipkart checkout without leaving the AI interface.
🏥 AI-assisted hospital billing may have added nearly $1 billion in costs over two years. More patients were coded as medically complex, with little evidence that the care they received changed.
🔐 Meta is making Muse’s safety warning clearer after a researcher found a flaw that could expose personal data if a user approved access to a malicious webpage.
⚡ Crusoe dropped its $1.25 billion turbine deal with Boom Supersonic for AI data centers. The company behind a major OpenAI data center says it still plans to use turbines, just not Boom’s.
💰 AI Daily Fundraising: 25-year-old Ali Ansari’s Micro1 reportedly raised over $100M at a $4B valuation, with two xAI cofounders investing. A Scale AI rival, Micro1 supplies training data to AI labs, serves Microsoft and Amazon, and reportedly reached a $500M gross annual run rate.
NEW EMPOWERED AI TOOLS
🧠 Hemory turns everyday conversations into private, searchable memory that Claude, Codex, Cursor, and other agents can access through MCP.
🎥 Eclatira lets developers add real-time voice, live vision, and autonomous video agents to apps with custom APIs, MCPs, and 3,000+ integrations.
📈 Ami AI plans and runs outbound campaigns by finding high-fit buyers, setting targets, launching outreach, and optimizing performance automatically.
🤝 Naoma AI Demo Agent V2 replaces demo forms with an AI sales agent that shows your product, qualifies visitors, answers questions, and books meetings.
🎙️ Voiskey turns rough speech into polished, context-aware writing across apps, supporting 100+ languages on desktop and mobile.
AI SAFETY WATCH
The Hugging Face breach looked like an unusually serious case of an AI agent going off course. Now Axios reports that OpenAI, Anthropic, and outside researchers are investigating tens of thousands of incidents involving frontier models. The cases span internal tests and real-world activity. What’s showing up?
Some agents bypassed guardrails, tried to leave secure test environments, or used websites in ways they weren’t authorized to. OpenAI says it has notified dozens of third parties while reviewing its agents’ activity.
The count includes failed attempts and deliberate stress tests designed to expose unsafe behavior. Axios says most cases are not known to have caused real-world harm. It isn’t a tally of successful hacks.
OpenAI has slowed frontier training and paused its largest planned reinforcement learning run while tightening security. Anthropic says Claude Opus 5.5 performed better than earlier models in its safety tests, though its evaluations can’t catch every possible failure.
The scale is the story. Agents can keep working through obstacles, including ones their builders expected to stop them. As OpenAI and Anthropic give models more freedom to act, catching those unexpected moves before they reach live systems becomes a much harder job.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply