Answer in brief
AI cybersecurity moved from chat assistants toward agentic systems in late summer 2026. On September 1, CrowdStrike and NVIDIA launched SafeMind, which pairs an offensive model, Red Tempest, with a defensive model, Blue Solano, built on NVIDIA Nemotron and trained on CrowdStrike data. In August, OpenAI released GPT-5.6 Cyber only to trusted partners such as Accenture, IBM, CrowdStrike and Cloudflare. CrowdStrike's performance figures are its own and have not been independently verified.
What CrowdStrike and NVIDIA launched at Fal.Con 2026
AI cybersecurity took a clear step from chat assistants toward autonomous systems on September 1, 2026. At its Fal.Con conference in Las Vegas, CrowdStrike launched SafeMind, an agentic system for defenders created with NVIDIA, together with two purpose-built security models: Red Tempest for offensive testing and Blue Solano for defense.
According to CrowdStrike, the models are built on NVIDIA's open Nemotron models and trained on its own data: Falcon sensor telemetry, threat intelligence, event annotations from its Falcon Complete managed detection and response service and fifteen years of incident-response fieldwork. Training and inference run on CoreWeave's AI cloud.
SafeMind runs natively inside the CrowdStrike Falcon platform. The company said standalone access to the models is available through its Project QuiltWorks program, and NVIDIA's write-up of the event described a new offering, Falcon IQ, as the way CrowdStrike puts that program into operation.
Red Tempest and Blue Solano: an attacker and a defender in one loop
The idea behind SafeMind is to put an offensive model and a defensive model in the same loop. Red Tempest behaves like a red team, looking for weaknesses in an environment, while Blue Solano works like a blue team that detects suspicious activity, investigates it and helps close the gaps it finds.
George Kurtz, CrowdStrike's founder and chief executive, described a system that finds weaknesses, strengthens protection and gets smarter with every cycle. In NVIDIA's account of the event, he framed the problem bluntly: attackers already had frontier AI, and defenders did not.
NVIDIA said Blue Solano is based on its Nemotron 3 Super model and that Falcon IQ uses more than 50 agents working together. The design reflects a wider shift in security software, away from one assistant answering questions and toward many specialized agents that each handle part of an investigation.
How to read CrowdStrike's 29%, six-times and 99% claims
CrowdStrike says its models deliver a 29% higher detection rate than leading frontier models and open-source baselines, six times faster end-to-end remediation and 99% cost savings on detection and remediation. NVIDIA's write-up described Blue Solano as reaching higher accuracy at 99% lower cost than leading frontier models.
These are figures from the companies that built and sell the systems. The announcements do not include an independent evaluation, and when detection tools are compared, the exact baseline models, the test data and the false-positive rate matter as much as the headline percentage.
A 99% cost saving describes the cost of detection and remediation with these models against the baselines CrowdStrike chose. It is not the same as a 99% cut in an organization's total security spending, which also covers staff, licenses, tooling and the work of responding to incidents.
Why speed dominates the pitch: 27-second breakouts
The case for agentic defense rests on speed. Citing CrowdStrike data, NVIDIA's account of the event reported an 89% rise in AI-enabled attacks over the past year and a fastest eCrime breakout time of just 27 seconds.
Breakout time is the interval between an intruder's first foothold on one machine and their move to other systems on the same network. When the fastest attackers can make that move in under half a minute, a process that depends on an analyst reading an alert, opening a ticket and escalating it is structurally too slow.
That is the gap vendors want AI agents to close: triaging alerts, gathering context and starting containment in seconds rather than minutes or hours. In the launch announcement, NVIDIA's Bartley Richardson summed up the promise as defenders being able to act at machine speed.
OpenAI's GPT-5.6 Cyber and access for trusted partners
OpenAI moved in a similar direction a few weeks earlier. On August 10, 2026, TechCrunch reported that the company had launched GPT-5.6 Cyber, a version of its GPT-5.6 model designed for defensive work such as security testing and vulnerability research.
The model is not generally available. OpenAI offers it only to trusted customer partners, a group that includes Accenture, IBM, CrowdStrike and Cloudflare, as part of an expanded Daybreak service with two tiers: Blue for basic services and Red for a more advanced toolkit.
TechCrunch framed the launch against a growing number of AI-led attacks, including cases in which AI agents were used for social engineering. The restricted rollout shows how seriously developers now treat the dual-use risk of models that are good at finding vulnerabilities.
Gated access is becoming the rule for security-grade models
Put side by side, the two launches share a pattern. CrowdStrike delivers its models inside its own platform or through a defined program, and OpenAI limits its cyber model to vetted partners. Neither company is offering its most capable security model as an open, anonymous download.
The reason is the dual-use problem. A model that can map an organization's weaknesses for its defenders could, in the wrong hands, map them for an intruder. Gating access by customer relationship, contract and use case is an attempt to keep the defensive benefit while slowing misuse.
CrowdStrike also appears on both sides of this story: as one of OpenAI's named partners for GPT-5.6 Cyber and as the builder of its own models with NVIDIA. For a large security vendor, drawing on more than one model source looks like a way to avoid depending on a single AI supplier.
What agentic defense changes inside a security operations center
In a typical security operations center, analysts spend much of their time on triage: deciding which alerts matter, pulling logs and reconstructing what happened. Agentic systems aim to take on that groundwork, so that an analyst starts from an assembled investigation rather than a raw alert.
The bigger change is remediation. Once a system can isolate a machine, reset credentials or stop a process on its own, the question shifts from how accurate its analysis is to which actions it is allowed to take without a person approving them.
That turns governance into a technical requirement. Clear permission levels, a log of every automated step and a quick way to reverse an action matter as much as detection quality, because a confident mistake made at machine speed can disrupt a business as surely as an attack.
Before an AI agent can act on your network: three checks
Start with the evidence. Ask any vendor which baselines its benchmarks used, how the test data relates to your environment and what the false-positive rate looks like, then run a trial on your own systems and past incidents before trusting a published percentage.
Then look at the data. Security telemetry reveals a great deal about an organization, so confirm where it is processed, how long it is kept and whether it is used to improve models that serve other customers.
Finally, define the human role. Decide which automated actions are allowed, who is accountable when an agent acts, and how staff will keep the skills to investigate incidents themselves if the AI is wrong, unavailable or targeted by attackers itself.
The next signals for SafeMind and GPT-5.6 Cyber
The first signal is independent evidence. Customer case studies, third-party evaluations and red-team results from outside the vendors will show whether the headline detection and cost figures hold up in real networks.
The second is access: how quickly the models reach customers beyond early programs, and whether OpenAI's partner-only approach widens or stays narrow as the first deployments are assessed.
The third is the attackers' response. Defenders are adopting agentic AI because attackers already use AI, and the real test of this generation of tools will be whether breakout times and successful intrusions fall as the tools spread.
Practical checklist
- Ask a vendor for the baseline, dataset and false-positive rate behind any detection or speed claim.
- Decide which actions an AI agent may take on its own and which need a named person to approve them.
- Confirm where your telemetry is processed and whether it is used to train models shared with other customers.
- Keep an audit trail of every automated containment or remediation step so it can be reviewed and reversed.
- Test an AI defender against your own environment and past incidents before relying on published benchmarks.
Questions and answers
What is CrowdStrike SafeMind?
An agentic security system that CrowdStrike launched with NVIDIA on September 1, 2026. It combines an offensive model, Red Tempest, with a defensive model, Blue Solano, and runs natively in the CrowdStrike Falcon platform.
What are Red Tempest and Blue Solano?
Red Tempest is a red-team model that probes an environment for weaknesses, and Blue Solano is a blue-team model focused on detection and response. Both are built on NVIDIA's open Nemotron models.
Is GPT-5.6 Cyber available to everyone?
No. TechCrunch reported in August that OpenAI offers GPT-5.6 Cyber only to trusted customer partners, including Accenture, IBM, CrowdStrike and Cloudflare, for security testing and vulnerability research.
Are the 29% and 99% figures independently verified?
No. They come from CrowdStrike's own comparisons with leading frontier and open-source models. Treat them as vendor claims until independent evaluations or customer results are published.
Will AI replace security analysts?
These launches do not claim that. The systems are pitched as speeding up detection, investigation and remediation, while decisions about risky automated actions still need human oversight and clear accountability.