OpenAI Unveils GPT-5.6 and ChatGPT Work: Agentic AI Boosts Cybersecurity Automation

OpenAI this week unveiled GPT-5.6 alongside ChatGPT Work, positioning its newest flagship as an agentic AI platform designed to complete extended, multi-step tasks across enterprise applications, files, web services, and desktop environments.

The release introduces three tiers, Sol (flagship), Terra, and Luna, and a shift toward persistent task execution that could materially change how security teams investigate, remediate, and automate operations.

ChatGPT Work reframes the assistant model as a persistent agent rather than a stateless conversational tool. According to OpenAI, the system can gather information from connected apps and organizational workflows, decompose complex objectives into smaller steps, and continue working for hours when required.

OpenAI Unveils GPT-5.6 and ChatGPT Work

Built-in Codex-derived capabilities enable the agent to interact across web, mobile, and desktop surfaces, allowing analysts to monitor progress, answer clarifying questions, redirect actions, and approve sensitive steps during execution.

For cybersecurity teams, that means a single agent might ingest threat feeds, correlate them with internal asset inventories, generate remediation trackers, build executive slide decks, and produce technical documentation, all while preserving context throughout the workflow.

A defining technical innovation in GPT-5.6 Sol is “ultra,” a high-capability mode that coordinates four agents in parallel by default and was evaluated in larger groups of up to 16 agents in selected benchmarks.

The multi-agent architecture permits parallelized workstreams, for example, concurrently reviewing vulnerable code, validating exploit conditions, testing patches, analyzing logs, and assembling a final report.

That capability addresses time-to-triage and time-to-remediation pressures faced by SOCs, vulnerability management programs, red teams, and application security operations, where simultaneous sub-tasks can significantly compress the investigation-to-mitigation window.

OpenAI reported significant gains in offensive-security benchmarks. On ExploitBench2, which measures the chain from vulnerable-code analysis to arbitrary code execution, GPT-5.6 Sol achieved 73.5% versus 47.9% for GPT-5.5 under similar token budgets.

In ExploitGym3, which evaluates turning real-world vulnerabilities into functioning exploits, GPT-5.6 recorded a 24.9% pass rate within two hours, rising to 33.7% at six hours. SEC-Bench Pro, focused on proof-of-concept generation against complex software, returned a 71.2% score for Sol, compared with 45.8% for the prior model.

OpenAI emphasized that GPT-5.6 is stronger at finding, reproducing, and fixing vulnerabilities than at autonomously conducting attacks against hardened environments.

To balance capability and risk, OpenAI pairs GPT-5.6 with its Daybreak Trusted Access for Cyber program and a layered safety stack. Access to expanded cyber functionality, authorized vulnerability validation, malware analysis, detection engineering, secure code review, and patch validation, requires qualified, verified users.

Real-time checks, reasoning-based monitors, account-level enforcement, hardware-backed passkeys, and restrictions on high-risk scenarios form the enforcement perimeter.

OpenAI also surfaced pricing and a model feature matrix: Sol as the top-tier research and complex cyber workloads model, Terra for balanced coding and security tasks at lower cost, and Luna as the fastest, cost-efficient tier for high-volume SOC automation.

Security practitioners see potential productivity benefits but warn of new operational challenges. Agentic workflows that can autonomously read, modify, and produce artifacts across enterprise systems magnify the impact of misconfigurations, over-permissive integrations, or compromised credentials.

Malicious actors could attempt to abuse similar agentic techniques if they gain access, while defenders must validate the AI’s findings and ensure robust approval gates for any automated remediation.

Enterprises adopting GPT-5.6 and ChatGPT Work should implement strict access controls, audit trails, human-in-the-loop checkpoints for high-risk actions, and staged rollouts for automation tasks.

When paired with careful governance, the agentic model promises to accelerate investigative cycles and lift repetitive workloads, but it also raises the bar for operational security disciplines in an era where AI can both accelerate fixes and, if misused, amplify harm.