AI Development

OpenAI shares a framework for tracking, investigating, and disclosing model misalignment, alongside six reports of unexpected or concerning model behavior.

Fyxer uses OpenAI models, fine-tuning, memory, and real user feedback to organize inboxes and draft emails in each user’s voice.

Perplexity uses Astra to write communications, change software, and monitor production systems, and checks in much less frequently than with earlier models.

<p><strong><a href="https://hugovk.dev/blog/2026/soft-deprecating-re.match/">Soft-deprecating re.match()</a></strong></p> Python has a concept of <a href="https://peps.python.org/pep-0387/#soft-deprecation">soft deprecation</a>, where APIs are marked as "should no longer be used to write new code" without any promise/threat to remove them in the future.</p> <p>Python 3.15 release manager Hugo van Kemenade describes how in the upcoming 3.15 release soft deprecation has come for the venerable but

<p>Graham Dumpleton's new monkey patching package <a href="https://wrapture.readthedocs.io/">wrapture</a> is shaping up to be an indispensable tool for Python developers. I'm not sure why I've seen so little buzz about it!</p> <p>Graham has been posting new tutorials for it almost daily since <a href="https://simonwillison.net/2026/Aug/31/introducing-wrapture/">the initial release</a> on August 31st. Here's everything he's published so far:</p> <ul> <li><a href="https://grahamdumpleton.me/posts/

GPT‑Live‑1 brings natural, full-duplex voice conversations to the API, with stronger instruction following, custom voices, and telephony support.

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini-3-8__fairwind-program__b.max-600x600.format-webp.webp">The Fairwind Program is a limited access program for governments and trusted partners to use our cyber defense tools.

OpenAI shares findings from the Hugging Face security incident and the steps we’re taking to strengthen AI model security, monitoring, and alignment.

Approved Daybreak partners can use OpenAI’s frontier cyber models to deliver authorized, governed cybersecurity services to customers.

Learn how GPT-5.6 powers Microsoft 365 Copilot with stronger AI capabilities across Word, Excel, PowerPoint, Chat, and Cowork for faster, higher-quality work.

OpenAI introduces Patch the Planet, a Daybreak initiative helping open-source maintainers find, validate, and fix vulnerabilities with AI and expert review.

OpenAI supports the EU Code of Practice on AI content transparency, advancing provenance standards and tools to help people understand AI-generated content.

See how LSEG uses OpenAI to scale trusted AI across its global business, accelerating insights, shrinking release cycles, and empowering 4,000 employees.

How Braintrust engineers use Codex with GPT-5.5 to run experiments and code faster.

Warp uses GPT-5.5 and OpenAI models to coordinate coding agents across local, cloud, and open-source development workflows.

OpenAI expands Trusted Access for Cyber with GPT-5.5 and GPT-5.5-Cyber, helping verified defenders accelerate vulnerability research and protect critical infrastructure.

Explore new realtime voice models in the OpenAI API that can reason, translate, and transcribe speech, enabling more natural and intelligent voice experiences.

Introducing Trusted Contact in ChatGPT, an optional safety feature that notifies someone you trust if serious self-harm concerns are detected.

Leading security firms and enterprises join OpenAI’s Trusted Access for Cyber, using GPT-5.4-Cyber and $10M in API grants to strengthen global cyber defense.

OpenAI updates the Agents SDK with native sandbox execution and a model-native harness, helping developers build secure, long-running agents across files and tools.

<p><strong><a href="https://github.com/simonw/datasette/pull/2689">datasette PR #2689: Replace token-based CSRF with Sec-Fetch-Site header protection</a></strong></p> Datasette has long protected against CSRF attacks using CSRF tokens, implemented using my <a href="https://github.com/simonw/asgi-csrf">asgi-csrf</a> Python library. These are something of a pain to work with - you need to scatter forms in templates with <code><input type="hidden" name="csrftoken" value="{{ csrftoken() }}"></code>

<p><strong><a href="https://openai.com/index/scaling-trusted-access-for-cyber-defense/">Trusted access for the next era of cyber defense</a></strong></p> OpenAI's answer to <a href="https://simonwillison.net/2026/Apr/7/project-glasswing/">Claude Mythos</a> appears to be a new model called GPT-5.4-Cyber:</p> <blockquote> <p>In preparation for increasingly more capable models from OpenAI over the next few months, we are fine-tuning our models specifically to enable defensive cybersecurity use case

OpenAI expands its Trusted Access for Cyber program, introducing GPT-5.4-Cyber to vetted defenders and strengthening safeguards as AI cybersecurity capabilities advance.

<p><strong>Research:</strong> <a href="https://github.com/simonw/research/tree/main/servo-crate-exploration#readme">Exploring the new `servo` crate</a></p> <p>In <a href="https://servo.org/blog/2026/04/13/servo-0.1.0-release/">Servo is now available on crates.io</a> the Servo team announced the initial release of the <a href="https://crates.io/crates/servo">servo</a> crate, which packages their browser engine as an embeddable library.</p> <p>I set Claude Code for web <a href="https://github.com/

<p>Thanks to a <a href="https://twitter.com/RahimNathwani/status/2039961945613209852">tip from Rahim Nathwani</a>, here's a <code>uv run</code> recipe for transcribing an audio file on macOS using the 10.28 GB <a href="https://huggingface.co/google/gemma-4-E2B">Gemma 4 E2B model</a> with MLX and <a href="https://github.com/Blaizzy/mlx-vlm">mlx-vlm</a>:</p> <pre><code>uv run --python 3.13 --with mlx_vlm --with torchvision --with gradio \ mlx_vlm.generate \ --model google/gemma-4-e2b-it \ --audio

<p><strong><a href="https://sqlite.org/releaselog/3_53_0.html">SQLite 3.53.0</a></strong></p> SQLite 3.52.0 was withdrawn so this is a pretty big release with a whole lot of accumulated user-facing and internal improvements. Some that stood out to me:</p> <ul> <li><code>ALTER TABLE</code> can now add and remove <code>NOT NULL</code> and <code>CHECK</code> constraints - I've previously used my own <a href="https://sqlite-utils.datasette.io/en/stable/python-api.html#changing-not-null-status">sqlit

Microsoft is starting to remove "unnecessary" Copilot buttons from its Windows 11 apps. In the latest version of the Notepad app for Windows Insiders, Microsoft has removed the Copilot button in favor of a "writing tools" menu. The Copilot button in the Snipping Tool app also no longer appears when you select an area to […]

April 10, 2026SecurityOur response to the Axios developer tool compromiseLoading…ShareWe recently identified a security issue involving a third-party developer tool, Axios, that was part of a widely reported, broader industry incident⁠(opens in a new window). Out of an abundance of caution we are taking steps to protect the process that certifies our macOS applications are legitimate OpenAI apps. We found no evidence that OpenAI user data was accessed, that our systems or intellectual property was compromised, or that our software was altered. We are updating our security certificates, which will require all macOS users to update their OpenAI apps to the latest versions. This helps prevent any risk—however unlikely—of someone attempting to distribute a fake app that appears to be from OpenAI. You can update safely through an in-app update or at the official links below:ChatGPT Desktop⁠(opens in a new window)Codex App⁠(opens in a new window)Codex CLI⁠(opens in a new window)Atlas⁠(opens in a new window)The security and privacy of your information are a top priority. We’re committed to being transparent and taking quick action when issues arise. We're sharing more technical details and FAQs below.What happened and what we are doingOn March 31, 2026 (UTC), Axios, a widely used third-party developer library, was compromised as part of a broader software supply chain attack.⁠(opens in a new window) At that time, a GitHub Actions workflow we use in the macOS app-signing process downloaded and executed a malicious version of Axios (version 1.14.1). This workflow had access to a certificate and notarization material used for signing macOS applications, including ChatGPT Desktop, Codex, Codex-cli, and Atlas. This certificate helps customers know that software comes from the legitimate developer, OpenAI. Our analysis of the incident concluded that the signing certificate present in this workflow was likely not successfully exfiltrated by the malicious payload due to the timing of the payload execution, certificate injection into the job, sequencing of the job itself, and other mitigating factors. Nevertheless, out of an abundance of caution we are treating the certificate as compromised, and are revoking and rotating it. Effective May 8, 2026, older versions of our macOS desktop apps will no longer receive updates or support, and may not be functional. These versions represent the earliest releases signed with our updated certificate:ChatGPT Desktop: 1.2026.051Codex App: 26.406.40811Codex CLI: 0.119.0Atlas: 1.2026.84.2Investigation and remediation effortsAs part of our investigation and response, we engaged a third-party digital forensics and incident response firm, rotated our macOS code signing certificate, published new builds of all relevant macOS products with the new certificate, and are working with Apple to ensure software signed with the previous certificate cannot be newly notarized. We have also reviewed all notarization of software using our previous certificate to confirm no unexpected software notarization occurred with these keys, and validated that our published software did not have unauthorized modifications. At this time, we have found no evidence of compromise or risk to existing software installations.In the event that the certificate was successfully compromised by a malicious actor, they could use it to sign their own code, making it appear as legitimate OpenAI software. We have stopped new software notarizations using the old certificate, so new software signed with the old certificate by an unauthorized third party would be blocked by default by macOS security protections unless a user explicitly bypasses them. Once we fully revoke our certificate on May 8th, 2026, new downloads and launches of apps signed with the previous certificate will be blocked by macOS security protections.The root cause of this incident was a misconfiguration in the GitHub Actions workflow, which we have addressed. Specifically, the action in question used a floating tag, as opposed to a specific commit hash, and did not have a configured minimumReleaseAge for new packages. FAQWere OpenAI products or user data compromised?No. We have found no evidence that OpenAI products or user data were compromised or exposed.Have you seen malware signed as OpenAI?No. We have found no evidence that the potentially exposed notarization and code signing material have been misused, and we have confirmed all notarization events with the impacted material were expected. Do I need to change my password?No. Passwords and OpenAI API keys were not affected.Does this affect iOS, Android, Linux, or Windows?No. This only affects OpenAI macOS apps. This does not affect the web versions of our software. Why are you asking me to update my Mac apps?OpenAI identified exposure in a GitHub Actions workflow involved in the macOS app-signing process. Because the exposed workflow was related to macOS app signing, we are proactively rotating the notariza

<p><strong>Release:</strong> <a href="https://github.com/simonw/scan-for-secrets/releases/tag/0.3">scan-for-secrets 0.3</a></p> <blockquote> <ul> <li>New <code>-r/--redact</code> option which shows the list of matches, asks for confirmation and then replaces every match with <code>REDACTED</code>, taking escaping rules into account.</li> <li>New Python function <code>redact_file(file_path: str | Path, secrets: list[str], replacement: str = "REDACTED") -> int</code>.</li> </ul> </blockquote> <p>T

<p><strong>Release:</strong> <a href="https://github.com/simonw/scan-for-secrets/releases/tag/0.1">scan-for-secrets 0.1</a></p> <p>I like publishing transcripts of local Claude Code sessions using my <a href="https://github.com/simonw/claude-code-transcripts">claude-code-transcripts</a> tool but I'm often paranoid that one of my API keys or similar secrets might inadvertently be revealed in the detailed log files.</p> <p>I built this new Python scanning tool to help reassure me. You can feed it

<p><strong>Release:</strong> <a href="https://github.com/simonw/research-llm-apis/releases/tag/2026-04-04">research-llm-apis 2026-04-04</a></p> <p>I'm working on a <a href="https://github.com/simonw/llm/issues/1314">major change</a> to my LLM Python library and CLI tool. LLM provides an abstraction layer over hundreds of different LLMs from dozens of different vendors thanks to its plugin system, and some of those vendors have grown new features over the past year which LLM's abstraction layer c

GDDRHammer, GeForge and GPUBreach hammer GPU memory in ways that hijack the CPU.

Mercor, a popular AI recruiting startup, has confirmed a security incident linked to a supply chain attack involving the open source project LiteLLM. The AI startup told TechCrunch on Tuesday that it was “one of thousands of companies” affected by a recent compromise of LiteLLM’s project, which was linked to a hacking group called TeamPCP. Confirmation of the incident comes as extortion hacking group Lapsus$ claimed it had targeted Mercor and gained access to its data. It’s not immediately clear how the Lapsus$ gang obtained the stolen data from Mercor as part of TeamPCP’s cyberattack. Founded in 2023, Mercor works with companies, including OpenAI and Anthropic, to train AI models by contracting specialized domain experts such as scientists, doctors, and lawyers from markets, including India. The startup says it facilitates more than $2 million in daily payouts and was valued at $10 billion following a $350 million Series C round led by Felicis Ventures in October 2025. Mercor spokesperson Heidi Hagberg confirmed to TechCrunch that the company had “moved promptly” to contain and remediate the security incident. “We are conducting a thorough investigation supported by leading third-party forensics experts,” said Hagberg. “We will continue to communicate with our customers and contractors directly as appropriate and devote the resources necessary to resolving the matter as soon as possible.” Earlier, Lapsus$ claimed responsibility for the apparent data breach on its leak site and shared a sample of data allegedly taken from Mercor, which TechCrunch reviewed. The sample included material referencing Slack data and what appeared to be ticketing data, as well as two videos purportedly showing conversations between Mercor’s AI systems and contractors on its platform. Techcrunch event This Week Only: Up to $482 savings for Disrupt 2026 Offer ends April 10, 11:59 p.m. PTYour next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to secure these savings. This Week Only: Up to $482 savings for Disrupt 2026 Offer ends April 10, 11:59 p.m. PTYour next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to secure these savings. San Francisco, CA | October 13-15, 2026 REGISTER NOW Hagberg declined to answer follow-up questions on whether the incident was connected to claims by Lapsus$, or whether any customer or contractor data had been accessed, exfiltrated, or misused. The compromise of LiteLLM originally surfaced last week after malicious code was discovered in a package associated with the Y Combinator-backed startup’s open source project. While the malicious code was identified and removed within hours, the incident drew scrutiny due to LiteLLM’s widespread use around the internet, with the library downloaded millions of times per day, per security firm Snyk. The incident also prompted LiteLLM to make changes to its compliance processes, including shifting from controversial startup Delve to Vanta for compliance certifications. It remains unclear how many companies were affected by the LiteLLM-related incident or whether any data exposure occurred, as investigations continue.

<p><strong><a href="https://ngrok.com/blog/quantization">Quantization from the ground up</a></strong></p> Sam Rose continues <a href="https://simonwillison.net/tags/sam-rose/">his streak</a> of publishing spectacularly informative interactive essays, this time explaining how quantization of Large Language Models works (which he says might be "<a href="https://twitter.com/samwhoo/status/2036845101561835968">the best post I've ever made</a>".)</p> <p>Also included is the best visual explanation I'

Relatively light at just 2 billion parameters, the model is meant for use with consumer-grade GPUs for those who want to self-host it. It currently supports 14 languages.

Listen to the session or watch below Whether it’s the race to find life on Mars, the campaign to outsmart killer asteroids, or the quest to make the moon a permanent home to astronauts, scientists’ efforts in space can tell us more about where humanity is headed. This subscriber-only discussion examines the progress and possibilities…

The fundraise includes $1 billion for investing in early-stage startups, and $2.5 billion for late-stage growth businesses.

Rather than working from scratch to figure out how to make AI safer for teens, developers can use these policies to fortify what they build.

<p><a href="https://marcelotryle.com/blog/2026/03/22/starlette-10-is-here/">Starlette 1.0 is out</a>! This is a really big deal. I think Starlette may be the Python framework with the most usage compared to its relatively low brand recognition because Starlette is the foundation of <a href="https://fastapi.tiangolo.com/">FastAPI</a>, which has attracted a huge amount of buzz that seems to have overshadowed Starlette itself.</p> <p>Kim Christie started working on Starlette in 2018 and it quickly

Building on top of a Chinese model feels particularly fraught right now.

Article URL: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646 Comments URL: https://news.ycombinator.com/item?id=47467913 Points: 191 # Comments: 115

Article URL: https://opencode.ai/ Comments URL: https://news.ycombinator.com/item?id=47460525 Points: 887 # Comments: 412

The company is reducing Copilot entry points on Windows, starting with Photos, Widgets, Notepad, and other apps.

Back to Articles Build a Domain-Specific Embedding Model in Under a Day Enterprise + Article Published March 20, 2026 Upvote 9 +3 Steve H steve-nvidia Follow nvidia Rucha Apte ruchaa01 Follow nvidia Sean Sodha ssodha-nv Follow nvidia Oliver Holworthy nvidia-oliver-holworthy Follow nvidia If you are building a RAG (Retrieval-Augmented Generation) system, you have likely hit this wall: Everything works… until it doesn’t. General-purpose embedding models are trained to understand the internet; not your contracts, manufacturing logs, proprietary chemical formulations or internal taxonomy. They capture broad semantic similarity, but they do not understand the fine-grained distinctions that matter in your domain. Fine-tuning an embedding model can improve the performance of your retrieval pipeline when off-the-shelf models fail to effectively capture domain-specific nuances. Despite how critical embeddings are to RAG performance, the process remains surprisingly fragmented, the skills required are specialized, and the time investment is daunting. With a single GPU and less than a day of training time, you can transform a general-purpose embedding model into one that truly understands your domain, no manual labeling required. To help you hit the ground running, we are also releasing a ready-to-use synthetic training dataset generated from NVIDIA's public documentation using this exact pipeline. Using this data and the recipe, we saw over 10% improvement in both Recall@10 and NDCG@10. Atlassian applied this recipe to fine-tune on their JIRA dataset, increasing Recall@60 from 0.751 to 0.951, a 26% improvement - on a single GPU. 🔗Quick Links to Dataset and Codes: Embedding Model GitHub Synthetic dataset on NVIDIA’s public documents 🧑‍💻Open Source Projects Recipe Integrates: NeMo Data Designer for synthetic data generation NeMo Automodel for embedding model training BEIR for Information retrieval evaluation NeMo Export-Deploy for ONNX/TensorRT conversion NVIDIA NIM for production inference serving 📋Prerequisites: A directory of domain documents (text files - .txt, .md, or similar) A valid NVIDIA API key (free at build.nvidia.com) NVIDIA Ampere GPU or newer with at least 80GB memory (with Compute Capability >= 8.0) This tutorial has been tested on 1xA100 (80GB), and 1xH100 (80GB) By the end of this post, you’ll know how to answer:📄 Generate training data from domain documents without labeled data🎯 Use hard negative mining for effective contrastive training🔗 Improve embedding quality with multi-hop queries⚙️ Fine-tune a bi-encoder embedding model📊 Evaluate whether fine-tuning improves retrieval🚀 Deploy the fine-tuned model in your pipeline ⚙️Setup In this tutorial, we will finetune the base model Llama-Nemotron-Embed-1B-v2 - a 1-billion-parameter embedding model that balances quality and inference cost. To get started, follow this setup guide. 📚 Step 1: Generate Training Data from Documents Fine-tuning an embedding model requires thousands of (query, relevant document) pairs. Most use cases don’t have this data readily available. Creating it manually is expensive, slow, and often biased by the annotator’s personal interpretation of what’s “relevant.”Instead of labeling data by hand, you can use an LLM (nvidia/nemotron-3-nano-30b-a3b) to read your documents and automatically generate high-quality synthetic question–answer pairs. nemotron embed sdg -c default corpus_dir=./data/my_domain_docs How does it work? Behind the scenes, this runs a four-stage synthetic data generation (SDG) pipeline powered by NeMo Data Designer: What does the output look like? Source document chunk: The thermal design power (TDP) of the H100 GPU is 700W in SXM form factor. The cooling solution must maintain junction temperature below 83°C under sustained workloads. Liquid cooling is recommended for dense deployments exceeding 4 GPUs per node, as air cooling cannot dissipate sufficient heat in standard 2U chassis configurations. Generated QA pairs: { "question": "What cooling approach is recommended when deploying more than 4 H100 GPUs per server node?", "answer": "Liquid cooling is recommended for dense deployments exceeding 4 GPUs per node, as air cooling cannot dissipate sufficient heat in standard 2U chassis configurations.", "query_type": "contextual", "reasoning_type": "factual", "question_complexity": 3, "segment_ids": [1], "quality_score": 8.5 } { "question": "How does the 700W TDP of the H100 SXM constrain the choice between air and liquid cooling in multi-GPU configurations?", "answer": "The 700W TDP generates substantial heat that must be dissipated to keep junction temperatures below 83°C. In dense configurations exceeding 4 GPUs per node, air cooling in standard 2U chassis cannot handle this thermal load, making liquid cooling necessary.", "query_type": "multi_hop", "reasoning_type": "causal", "question_complexity": 4, "segment_ids": [1, 2], "hop_count": 2, "quality_score": 9.0 } Notice the difference: the first question is

Trump’s AI framework pushes federal preemption of state laws, emphasizes innovation, and shifts responsibility for child safety toward parents while laying out lighter-touch rules for tech companies.

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/25367___BRS___Aspen_Security_Fo.max-600x600.format-webp_J9uPoFt.webp">Google is making new investments, building new tools and developing code security to improve open source security.

OpenAI introduces Trusted Access for Cyber, a trust-based framework that expands access to frontier cyber capabilities while strengthening safeguards against misuse.

More than one million customers around the world now use OpenAI to empower their teams and unlock new opportunities. This post highlights how companies like PayPal, Virgin Atlantic, BBVA, Cisco, Moderna, and Canva are transforming the way work gets done with AI.

JetBrains is integrating GPT-5 across its coding tools, helping millions of developers design, reason, and build software faster.