标签：#openai

In a review triggered by OpenAI’s Hugging Face incident, Anthropic discovered three of its AI models had breached real organizations during third-party evaluations.

After OpenAI's models broke into Hugging Face, Anthropic checked its own history and found three similar incidents

OpenAI has formally outlined a national science initiative designed to connect frontier AI models with government research infrastructure, National Laboratories, universities, and working scientists. The program is not a single model launch. Instead, it combines funded access, early product access, scientific campaigns, and an emphasis on fitting advanced AI into real research workflows. The initiative gives concrete form to OpenAI’s stated goal of helping scientists use increasingly capable models to accelerate discovery. In its official announcement on advancing the next era of national science , published July 22, 2026, the company describes a long-term strategy built around the U.S. Department of Energy’s Genesis Mission and collaborations with National Laboratories. The core proposition is that AI can contribute to hypothesis testing, simulations, and experimental work when it is deployed alongside scientific infrastructure and human expertise. That framing matters. OpenAI is positioning frontier models as tools that researchers direct and evaluate, rather than as a replacement for the institutions and specialists responsible for scientific work. What OpenAI is providing to scientific researchers OpenAI’s commitments span several types of access, from coding support for a broad research community to model capabilities and API funding for large campaigns. The announced provisions include: $4 million in Codex access for approximately 2,000 Genesis researchers at national labs and universities. $3 million in API support for two large scientific campaigns. Up to $10 million in API usage for participating researchers who reach a $2.5 million spending threshold. Access to GPT-Rosalind’s bioscience capabilities for national-lab researchers. Early access to selected models and features for trusted national-lab leaders preparing workflows and evaluations. Expanded access to advanced cyber capabilities for national-lab cybersecurity researchers. These commitments indicat

OpenAI has published a post-mortem examining an unusual pattern in its model testing: recurring references to “goblins” and “gremlins” in model outputs. The company’s official post, “Where the goblins came from” , published on April 29, 2026, frames the behavior as an emergent effect of reinforcement learning and human-feedback dynamics, not as a new product feature. Its practical message is more consequential than the metaphor suggests: unexpected model personas can affect the consistency, safety, and reliability that developers expect from AI systems. The published analysis provides the substantive context behind recent attention to a purported “goblin-level” post. Rather than indicating a model launch, OpenAI’s account suggests a narrower but important lesson about how optimization signals can inadvertently reinforce patterns in language models. For organizations using LLMs in production, the relevant question is not whether goblin-like language is amusing. It is whether teams can detect and address unexpected behaviors before those behaviors influence customer-facing, operational, or high-stakes workflows. What OpenAI documented OpenAI said the “goblin” and “gremlin” metaphors appeared during GPT-5.x testing and RLHF training. The company reported a notable increase in goblin-like language during GPT-5.5 testing when Codex was being evaluated. According to the post, the pattern emerged from reward-signal dynamics : persona-like responses were inadvertently reinforced through reinforcement learning and human feedback. That distinction matters. OpenAI does not characterize goblin behavior as a fixed capability or intentional model identity. It describes it as a byproduct that can arise at scale when a training and feedback process favors certain output patterns. The episode is therefore best understood as an alignment and evaluation lesson, rather than evidence of a separate “goblin” model, feature, or policy release. OpenAI also described a mitigation introduced

I am not by any means an expert at finance but I think I do now have some advice for people who are: Do not name your hedge fund anything that will be hilarious if it blows up. Don't use a name like "Long-Term Capital Management," or "Amaranth Advisors" (named for the floral symbol for […]

If the generative AI giant had followed well-known security best practices, it’s likely that its AI agent would never have escaped to the open internet and hacked multiple companies.

Picture a moment every person who runs an AI agent knows. A task is halfway done and starting to go wrong. The agent took a weird turn a few steps back and now it is confidently heading somewhere bad. You have to decide fast on this. Do you step in? And if you do a quick "wait, check your work" nudge will that actually fix it? Or do nothing? Or worse knock a run that was about to recover on its own off the rails? That question is the whole project. Here is the honest short version of what I found. Detecting a failure is not fixing it A lot of recent agent research is about failure attribution — figuring out which step in a long run broke everything. Useful but it stops one step short of what you need when you are on call. Knowing where it broke is not the same as knowing what to do about it . So I asked a blunter question: given a failure, which fix actually recovers the run and which ones quietly make it worse? To answer it without fooling myself I rewind each failing run to the exact step where it went wrong, apply one fix, let it play forward and check the real answer against a hard ground truth no LLM grading another LLM. And I always compare against a "do nothing" control, because some runs recover on their own, and I did not want to give my fixes credit for that (or miss a "fix" that's actually worse than leaving the agent alone). What a capable agent actually gets wrong First surprise: a decent agent mostly doesn't fail in the dramatic ways people worry about. It rarely loops, rarely forgets to answer, rarely fumbles a tool that throws an error in its face. It fails in two quieter ways and both are the same underlying mistake: acting on the surface of the situation instead of the real thing underneath. It makes up an answer it could have looked up. The fact it needs is sitting right there behind a tool call it just never makes, so it fills the gap with something plausible. Reads "manager: #202," never looks up who #202 is, asserts a name anyway. It trusts a t

When Microsoft reported killer fourth-quarter earnings for its fiscal 2026 year (which ended June 30), it tucked in an interesting little tidbit about how its investments in the two biggest, and competing, AI labs are doing.

Weng previously served as the VP of AI Safety Research at OpenAI.

OpenAI has introduced Codex Security in research preview, positioning it as a project-contextual security capability within the Codex ecosystem. The feature is designed to build threat models for a codebase, validate potential findings in sandboxed environments, and surface proposed patches for human review. The release is significant for teams evaluating AI-assisted application security because it frames security work as more than a static scan. OpenAI's current approach combines codebase context, validation, and reviewer oversight, while keeping the feature tied to its cloud and web product experience rather than presenting it as a standalone open-source security command-line tool. What OpenAI has released According to OpenAI's Codex Security research preview announcement , Codex Security is being introduced as a research-preview feature for customers using Codex web. OpenAI says the tool develops a threat model based on project context, investigates potential issues, validates findings in sandboxed environments, and presents patches for human review. That sequence matters because security tooling can produce findings that require substantial engineering effort to triage. A system that can incorporate project context and test a potential issue before escalating it is intended to make findings more actionable. However, OpenAI's own description still places a human reviewer in the decision loop for proposed changes. OpenAI says Codex Security is rolling out through Codex web for ChatGPT Pro, Enterprise, Business, and Edu customers. It is therefore an enterprise-oriented, cloud-hosted preview capability with controlled availability, not a generally available security product for every developer environment. The currently described workflow includes: Building a threat model using the context of a software project. Investigating and validating potential security findings in sandboxed environments. Surfacing proposed patches for human review. Providing access through Co

I watched a new tool try to get around the model safeguards of four major frontier companies. You might be surprised by how they performed.

In an interview with our friend Joanna Stern on her YouTube channel, OpenAI president Greg Brockman said the company is working on a "family of devices" for interacting with its AI models. However, Brockman didn't confirm reports that one of those devices is a smart speaker OpenAI's rumored to be launching in 2027, or earlier […]

The AI agent that escaped from OpenAI and hacked developer platform Hugging Face attacked other companies as well, OpenAI revealed on Tuesday. The update substantially widens the scope of an already concerning incident, which has alarmed industry insiders and fueled growing calls for stronger oversight on frontier AI systems. In an update to a blog […]

Earlier this month, OpenAI gave several of its AI models a task: complete a test designed to measure their cybersecurity capabilities. It put the systems in a sandboxed environment without an internet connection and set them off to work. What happened next is almost laughably silly - but also, as Adam Gleave, cofounder and CEO […]

In a new disclosure, OpenAI says its agent used exposed logins to gain access to at least four “publicly available services” in its unhinged quest to solve a test.

It's earnings season, and investors got an unpleasant surprise from Google: an increase on its spending estimate, to as much as $205 billion - from the last quarter's projection of up to $190 billion. Even the lower end of Google's new projected range - $195 billion - is much more than the company had previously […]

Published on Aug 18th, 2025 The Setup: When Everything Seems Perfect After successfully implementing a chatbot based on ChatGPT in my portfolio (as detailed in my previous "Redesign Portfolio" post), I was feeling pretty good about myself. The AI integration was working smoothly, users could ask questions about my skills and projects, and I had successfully created content embeddings that made the chatbot intelligent and contextually aware. Little did I know that this "completed" project was about to become the perfect candidate for my "Unfinished Projects" series. The Crisis: When OpenAI Goes Silent It started with a simple error message: "OpenAI API account suspended." At first, I thought it was a simple configuration issue. Maybe I had accidentally exposed my API key or hit some rate limit. But after checking my environment variables and account status, I realized the problem was deeper. My OpenAI account was suspended, and suddenly, my "completed" AI chat functionality was completely broken. Suddenly, my SmartNotes app which also powers the chat functionality went offline. This was supposed to be a finished project. Instead, it had become the latest addition to my stack of unfinished work. The Panic: Scrambling for Solutions The immediate challenge was clear: either abandon the chat functionality entirely or find an alternative solution. Given that I had already invested significant time in building the user interface and database integration, abandoning it wasn't an option. I started researching alternatives: Claude API : Limited availability and different pricing structure Local AI models : Required significant computational resources Hugging Face : Promising, but I had no experience with their inference API The clock was ticking, and my portfolio was broken. The Discovery: Hugging Face to the Rescue After several hours of research and testing, I discovered that Hugging Face offered inference APIs that could potentially replace OpenAI's functionality. The catc

Published on Aug 18, 2025 A New Era of AI-Powered Coding Begins I have installed Cursor on my laptop this weekend, and I am amazed at how much it speeds up my coding. I have a new debugging buddy!! This week, I have made several updates to the Portfolio website. The Challenge: When OpenAI Falls Short In my previous post, I shared the excitement of implementing a chatbot based on ChatGPT for my portfolio website. The initial experience was promising - I successfully created content embeddings and integrated them with OpenAI's API. However, as many developers know, relying on a single service provider can lead to unexpected roadblocks. When my OpenAI account encountered issues, I faced a critical decision: abandon the chat functionality or find an alternative solution. I chose the latter, embarking on a journey that would transform my portfolio's AI capabilities and teach me valuable lessons about building robust, fallback-ready systems. The Migration: Embracing Open Source AI The transition from OpenAI to Hugging Face wasn't just a simple API swap - it was a complete architectural evolution. Here's what I learned: 1. Model Selection Complexity Finding the right model on Hugging Face proved more challenging than expected. After testing several options: microsoft/DialoGPT-medium - No inference provider available gpt2 and distilgpt2 - Limited conversational capabilities Qwen/Qwen3-4B - Perfect fit with the nebius provider 2. Database Architecture Evolution The migration also prompted a database upgrade from MongoDB to Neon PostgreSQL. This wasn't just about changing providers - it was about building a more scalable, production-ready foundation for my portfolio. Technical Implementation: Building Resilience Streaming Responses for Better UX One of the most significant improvements was implementing streaming text responses. Instead of waiting for complete AI responses, users now see text appear word-by-word, creating a ChatGPT-like experience: // Streaming implementation

OpenAI's Hugging Face breach has reignited debate over AI alignment and control, exposing competing views on whether increasingly capable AI should be better aligned, better contained, or both.

After two years in stealth, Safe Superintelligence has announced a long-term partnership with Nvidia as it prepares to scale to its next phase.