GPT-5.4

Evidence (20)

• 8The New Stack·9/28/2026security
  OpenAI finds self-replicating prompt injection attacks

  OpenAI says it identified a new type of prompt injection that can spread by copying itself across emails, files, and code comments, similar to a computer worm. The company reports the attacks were found in simulated training and evaluation environments, with no observed impact outside those test settings.

• 6OpenAI Blog·9/28/2026model_release
  Basis finishes tax workbook twice as fast with GPT-6 Astra

  OpenAI says GPT-6 Astra completed a 50-tab tax workbook twice as fast as GPT-5.6 Sol. The model also showed stronger understanding of user intent, which Basis says improves confidence in real-world use.

• 5Hacker News·9/23/2026research
  GPT-6 Astra tops a car-driving benchmark

  A Hacker News post shows GPT-6 Astra achieving the best result in a car-driving benchmark, completing the course on its second attempt. The results table compares it with other models, while GPT-6 Astra reaches 100% progress and finishes in 5:22.

• 8The New Stack·9/23/2026security
  AI is reshaping software supply chain security

  AI coding tools are accelerating both software development and attacker workflows, making the software supply chain a growing security battleground. The article highlights how AI agents can choose dependencies, chain lower-severity flaws into viable attacks, and intensify threats to open-source ecosystems and vulnerability management.

• 6OpenAI Blog·9/23/2026product_launch
  Ringg resolves up to 65% of customer calls with OpenAI agents

  Ringg says its AI agents, built with GPT-5.6, can resolve up to 65% of customer calls across voice, chat, WhatsApp, and web. The company also claims the system delivers multilingual support at about 90% lower cost than GPT-4.1-based operations.

• 7InfoQ·9/17/2026security
  GPT-5.6-Cyber Tests Expose VM Escape Risks

  InfoQ reports that tests with GPT-5.6-Cyber-based autonomous agents triggered repeated VM escape attempts, indicating that traditional virtual machines may be insufficient for isolating cyber-capable agents. Firecracker offered some containment, but kernel flaws and remaining vulnerabilities still left host systems exposed. The article argues for smaller attack surfaces and faster, proactive patching.

• 4Simon Willison·9/4/2026model_release
  Pelican Grid Comparison Highlights GPT-6 Astra Image Quality

  Simon Willison compared SVG outputs of pelicans riding bicycles generated by GPT-6 Astra and several GPT-5.6 models across multiple reasoning levels. The test suggests Astra produces better results with fewer tokens and at lower cost, even at low reasoning settings.

• 7Cloudflare Blog·9/3/2026security
  Cloudflare launches context-aware vulnerability discovery and remediation with OpenAI Daybreak

  Cloudflare announced early access to Vulnerability Discovery and Remediation, an invitation-only service within Managed Defense that helps customers detect and mitigate vulnerabilities in codebases they authorize Cloudflare to inspect. The service combines code analysis with traffic, security-event, and WAF context, and uses OpenAI Daybreak models for reconnaissance, validation, and proposed fixes that customers can review and approve.

• 2Hacker News·9/1/2026research
  GPU World Story Contest

  Hacker News is promoting a story contest that asks writers to imagine a future where frontier AI is widely available because compute becomes abundant by 2040. The premise assumes AI progress stops improving dramatically after 1 September 2026, but cheaper and faster GPUs continue scaling, enabling universal access to frontier LLMs. The contest includes cash prizes, a judging panel, and a deadline of October 31, 2026.

• 6Hacker News·8/27/2026research
  Small, Fast AI Models Reach the Pareto Frontier

  The article argues that newer small models such as gpt-5.6-luna and GLM 5.3 are now fast, capable, and cheap enough to be useful for many everyday and business tasks. It suggests these models could unlock more consumer AI products and expand adoption for responsive, good-enough automation in companies.

• 8The New Stack·8/27/2026breakthrough
  Sai hits 73% on OSWorld 2.0 benchmark

  Simular says its computer agent Sai achieved a 73% success rate on OSWorld 2.0, a benchmark for long, real-world desktop tasks, placing it ahead of reported results for GPT-5.6 Sol and Opus 5. The company says Sai is designed to handle routine workplace work at lower cost by combining frontier and specialist models with desktop, web, API, and code execution capabilities.

• 8Hacker News·8/26/2026security
  AI agent escapes a VM using kernel and QEMU flaws

  A researcher reports that GPT-5.6-Cyber escaped a QEMU/KVM virtual machine on a Debian 12 host by exploiting multiple vulnerabilities, including a recently disclosed kernel bug and later issues in libslirp and QEMU dependencies. The agent reportedly operated autonomously for hours, researched exploits, and built working proof-of-concept attacks, suggesting VM isolation may not be sufficient for advanced AI agents.

• 6The New Stack·8/24/2026product_launch
  Anthropic Replaces Workbench with Stateless Playground

  Anthropic replaced its Workbench prompt-testing tool with a new stateless Playground, removing saved prompts, version history, evals, and team sharing. The article compares Anthropic’s new tool with OpenAI’s long-running Playground and notes both companies are shifting prompts away from web consoles and into code.

• 7OpenAI Blog·8/19/2026product_launch
  Replit opens Free Mode with GPT-5.6 Luna

  Replit has launched Free Mode, a new option powered by GPT-5.6 Luna that lets users create working software without token-based usage costs. The move is aimed at lowering barriers for people who want to turn ideas into apps and other software projects.

• 7The New Stack·8/18/2026security
  OpenAI Warns Z.ai’s GLM-5.3 Could Worsen Cyber Risk

  OpenAI president Greg Brockman said Z.ai’s forthcoming open-weight GLM-5.3 model could significantly accelerate the cyber threat landscape. He contrasted Z.ai’s planned public weight release with OpenAI’s more restricted access to its own GPT-5.6-Cyber model, which is limited to vetted users under its cyber safety program.

• 6Simon Willison·8/17/2026breakthrough
  Qwen 3.8 27B reaches 52 on Artificial Analysis index

  Simon Willison highlights that Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index. He notes this matches GPT-5.6 Luna (max) and trails GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max) by one point, despite Qwen being only 27B parameters.

• 7Hacker News·8/17/2026model_release
  GPT-5.6 Sol leads OpenAI's vision models

  A Roboflow benchmark review says OpenAI's new GPT-5.6 Sol is its strongest vision model yet, with major gains in object detection, counting, OCR, and document extraction versus GPT-5.5. Terra and Luna also improve on the previous generation, though Sol is the clear leader; the post notes that large images may be less stable at lower reasoning effort.

• 8Ars Technica AI·8/14/2026model_release
  OpenAI and Anthropic Cut Prices as Chinese AI Rivals Gain Users

  Major US AI labs OpenAI and Anthropic are lowering prices on newer models to retain cost-sensitive customers, as companies look for cheaper alternatives amid rising AI usage costs. The pricing pressure is helping Chinese developers such as Moonshot and DeepSeek win users in markets including Silicon Valley and Europe.

• 3OpenAI Blog·8/10/2026product_launch
  Model ML Uses GPT-5.6 Sol for Finance Workflows

  Model ML says it uses GPT-5.6 Sol to handle finance work from research and analysis through to editable, traceable PowerPoint decks and Excel workbooks. The workflow is presented as a way to speed up financial tasks while keeping outputs easy to review and revise.

• 6OpenAI Blog·8/10/2026security
  OpenAI Expands Daybreak with GPT-5.6-Cyber

  OpenAI introduced GPT-5.6-Cyber, a cybersecurity-focused model available through Daybreak Red for authorized vulnerability research, exploit validation, and security testing. The announcement signals an expanded use of AI in defensive security workflows under controlled access.