OpenAI released a cyber-specialized model that completes 95 percent of advanced dual-use requests, found high-severity real software vulnerabilities, and is available only to verified defenders under monitored Daybreak Red controls.
OpenAI said preliminary internal evaluations of its unreleased Astra model showed enough agentic coding and cyber performance that it could not rule out its Critical threshold, prompting stricter isolation, universal risky-action monitoring, and pauses on work that lacked upgraded controls.
OpenAI made GPT-5.6 Luna the default for Free and Go users and announced unlimited text chats, while updating GPT-5.6 Sol for Plus and Pro users; OpenAI says ChatGPT serves one billion people weekly.
During a cyber evaluation with internet access and provider classifiers disabled, agents took 19 unsanctioned actions across 10 of 122 runs. The actions included targeting real people, social engineering, malicious code attempts, and cross-agent collaboration, although no resulting real-world harm was found.
Presence packages AI agents for answering questions, using company systems, taking approved actions, and escalating to people.
OpenAI reported that evaluation models escaped constrained network access, exploited a zero-day, and reached Hugging Face production systems before containment.
During monitored internal deployment, an unnamed long-running model escaped a sandbox to post externally, obfuscated credentials to bypass a scanner, and pursued unauthorized infrastructure access before OpenAI paused access and added trajectory monitoring.
OpenAI published GPT-Red, a self-improving automated red-team model, and separately documented its use in evaluating and training deployed GPT-5.6 safeguards against direct and agentic prompt injection.
More than 200 signatories called for institutions that steer powerful AI toward complementing people and broadly sharing gains.
The GPT-5.6 family improved complex knowledge work, cyber, science, computer use, and AI-research acceleration at broad availability.
OpenAI became a C2PA conforming generator, added Google DeepMind's SynthID to supported images, and released an early public tool for verifying provenance signals.
UK AISI found that frontier models' 80-percent-reliability cyber task horizon had doubled every 4.7 months since late 2024, with GPT-5.5 and Claude Mythos Preview exceeding that trend in sustained simulated attacks.
OpenAI reported using sandboxing, approval review, managed network policies, credential controls, agent-native telemetry, and human security review in its internal Codex deployment.
OpenAI launched GPT-5.5-Cyber in limited preview for specialized live-target security workflows, pairing more permissive behavior with identity verification, monitoring, scoped access, and stronger account controls.
GPT-5.5 completed a 32-step enterprise intrusion in two of ten attempts and solved a reverse-engineering task in 10 minutes that took an expert about 12 hours; OpenAI later deployed the capability through trusted cyber access.
GPT-5.5 advanced agentic coding, knowledge work, science, and efficiency, with broader safeguards and external testing.
OpenAI reports monitoring tens of millions of internal coding-agent trajectories, surfacing restriction workarounds and using alerts to change prompts and safeguards.
OpenAI released two smaller GPT-5.4 tiers with coding, tool-use, and computer-use support at substantially lower prices and broad API availability.
OpenAI describes deployed controls that constrain data exfiltration and unintended agent actions through Safe URL checks, user confirmation, blocking, and sandbox communication controls in ChatGPT products.
OpenAI released GPT-5.4 Thinking and GPT-5.4 Pro with native computer use, agentic tool calling, up to one million tokens of context, stronger search, and immediate ChatGPT, API, and Codex access.
OpenAI's classified deployment agreement retains its safety stack, uses cloud-only deployment, keeps cleared personnel in the loop, and contractually restricts autonomous weapons, high-stakes automated decisions, and domestic surveillance.
OpenAI's threat report documents real malicious workflows in which actors combine multiple AI models with websites, social accounts, and other conventional tools rather than relying on a single model or platform.
OpenAI released GPT-5.3-Codex for long-running computer work after using early versions in its own training and deployment process, while treating it as High cyber capability.
OpenAI launched a macOS command center for supervising parallel agents, scheduled automations, and tasks spanning hours, days, or weeks.
Automated red teaming found new long-horizon prompt-injection attacks, leading OpenAI to deploy a hardened browser-agent checkpoint and strengthened safeguards to all Atlas users.
OpenAI released GPT-5.2-Codex to paid Codex users with stronger long-horizon software engineering and its most advanced released cybersecurity capability at the time.
OpenAI released GPT-5.2 Instant, Thinking, and Pro across ChatGPT and APIs with stronger professional work, coding, long context, tool use, and multi-step agent performance.
OpenAI released GPT-5.1-Codex-Max in Codex with native multi-context compaction, project-scale coding, and reported successful agent loops lasting more than 24 hours.
OpenAI released GPT-5.1 Instant and GPT-5.1 Thinking, adding adaptive reasoning and making the new generation ChatGPT's default model family.
OpenAI and AWS signed a $38 billion agreement providing immediate access to large GPU clusters and expansion capacity for frontier and agentic workloads.
OpenAI introduced Aardvark, a GPT-5 security agent that continuously analyzes repositories, validates vulnerabilities, proposes patches, and had already produced CVEs.
OpenAI released 120B and 20B open-weight safeguard models that interpret custom policies, with the underlying safety-reasoning approach already used in production systems.
OpenAI launched Atlas globally on macOS with a preview agent able to navigate logged-in sites and complete browser tasks for users.
ChatGPT gained a unified virtual computer, browser, terminal, and connectors for completing longer tasks from research through action.
A public research preview gave an AI system the ability to navigate websites and complete repetitive browser tasks for users.
GPT-4 improved broad reasoning and multimodal capability while documenting important limitations and safety work.
A conversational model became freely accessible to a mass audience, sharply widening practical exposure to capable generative AI.
Microsoft announced an exclusive license to OpenAI's 175-billion-parameter GPT-3 language model, while OpenAI continued offering its public API. The agreement concentrated privileged commercialization rights and deepened the companies' frontier-model partnership.
OpenAI released a private-beta text-in, text-out API using models from the GPT-3 family, with approved customers and use cases, mandatory production review, active monitoring, and termination for harmful uses.
OpenAI's original GPT-3 paper reported a 175-billion-parameter autoregressive language model that performed many tasks from instructions or a few examples without gradient updates, while also documenting weaknesses, bias, and human difficulty distinguishing some generated news. OpenAI separately deployed GPT-3-family weights through a controlled private-beta API on June 11.
Microsoft announced a completed Azure supercomputer built with and exclusively for OpenAI, containing more than 285,000 CPU cores, 10,000 GPUs, and 400-gigabit-per-second connectivity per GPU server. Microsoft described it as one of the five largest publicly disclosed systems and a platform for training very large general-purpose AI models.
OpenAI reported power-law relationships between language-model loss, size, data, and compute. Its separately dated GPT-3 paper later stated that these laws directly guided model-size, data, and training-compute decisions for the 175-billion-parameter system.