Introducing GhostWorks: A Practical AI Initiative from SpecterOps 

GhostWorks is an AI-focused engineering and research initiative at SpecterOps, focused on the disciplined exploration of frontier AI-enabled cybersecurity tooling. It is not a company-wide AI mandate, a replacement for product teams, or a promise that every workflow needs an agent. It exists to continue exploration of frontier AI-enabled tooling with pragmatism and discipline. We test emerging capabilities against real identity-security problems, document the limits, and hand off work that is ready to survive outside the lab. 

Putting AI in perspective 

LLM tools have entered cybersecurity operations with a variety of applications; everybody’s read about Mythos and Project Glasswing, for example. However, LLM tools aren’t just used for adversary tradecraft. LLMs are finding their way into every corner of our lives. Companies like Google, Apple, and Microsoft are pushing hard to make the average person want to reach for their phone to accomplish everyday tasks via speech and a phone camera. The scary or exciting part (depending on who you ask) is some of the demonstrations are compelling, but the universal use cases often feel unclear. 

When a new technology hits the industry and moves fast, with security and privacy concerns this serious, we’re all kind of stumbling in the dark. And yet, that’s not how people, or at least the vocal denizens of the internet, frame it. The internet offers two extremes: performance is 10x, and every CTO not engaging with LLM tools is falling behind, or any claimed improvement is immeasurable at-best or an outright lie at-worst. It’s either abundant optimism or dogged pessimism. But the real use-case, the effective application of a tool, always lives between the lines. There exists a very thin middle ground right now, an island of pragmatism. 

A tradecraft approach to AI 

Adversary tradecraft isn’t just about the adversarial context that these tools are operated in. Understanding tradecraft is the first step in being able to defend against it. LLM tooling, including Codex, GPT Cyber, Mythos, and Project Glasswing, is no different. There is a middle ground to be found here, and that middle ground is forged through pragmatic evaluation and strategic application. Big words for a simple outcome: find where the technology fits best, measure everything, and make informed decisions.  

This is at the heart of why we built GhostWorks. 

GhostWorks is about codifying the approach that SpecterOps has always taken when new adversary tradecraft, technologies, or methodologies appear with the momentum to stay. If attackers are creating exploits in hours, then it’s our job to figure out how they’re doing it and then to build tools to defend against it. 

GhostWorks is not the starting line for SpecterOps’s AI work. Before this initiative had a name, SpecterOps was already partnering with OpenAI on continuous adversarial red teaming and advanced skills training. SpecterOps is an active participant in OpenAI’s Trusted Access for Cyber (TAC) program. SpecterOps also helped build The Last Ones, a realistic enterprise cyber range used by the UK AI Security Institute (AISI) to measure frontier models against an intrusion path. GhostWorks gives that work an intentional home: a way to turn individual experiments, evaluations, and prototype tooling into disciplined engineering outcomes. 

The Ghostworks initiative exists to prototype AI-enabled capabilities. We test frontier capabilites against real identity-security problems and decide what should become product features, research initiatives, open source or internal tooling, or a documented failure. Some experiments will become demos. Some will become evaluation frameworks. Some will become lessons about where this technology fails. The point is to turn uncertainty into evidence and measurable results, not chase novelty. 

The engineering-focused approach we have chosen to take is one that has served us well in building software like Mythic, Ghostwriter, Nemesis, BloodHound and BloodHound Enterprise. Each of these efforts exists as an intersection of expert engineering and deep security domain expertise. We want to look at this not just as security practitioners, but as engineers, as builders, as people who are passionate about taking apart systems, understanding how they work, and then reconfiguring them to our purposes. 

What comes next 

In the coming weeks and months, we will be publishing the results of this research and testing through a series of blogs, and in discussions on the Know Your Adversary podcast. The purpose of GhostWorks is to facilitate a practical conversation for both defenders and red teamers about what this technology means, what it doesn’t, and how it applies in a real-world environment.  

You can read the first of these blogs here: Prompt Engineering for Security Agents: A Measurable Approach with GEPA

Stay tuned.  

Further reading 

• OpenAI: Security on the Path to AGI – OpenAI’s March 2025 security update names SpecterOps as a partner for continuous adversarial red teaming and advanced skills training. 

• SpecterOps: SpecterOps Selected for OpenAI’s Trusted Access for Cyber Program – SpecterOps’ April 2026 announcement explains how the Trusted Access for Cyber program supports research into frontier AI, chained permissions, lateral movement, and privilege escalation across identity platforms. 

• UK AISI: Our evaluation of OpenAI’s GPT-5.5 cyber capabilities – UK AISI’s April 2026 evaluation describes “The Last Ones,” a 32-step simulated corporate network attack range built with SpecterOps and used to evaluate model performance on end-to-end cyber tasks.