The narrative surrounding the "Mythos" model suggests it has single-handedly invented automated penetration testing. It hasn't. While the industry buzz treats this specific model as the genesis of AI-driven hacking, the reality is that Security Arsenal has been leveraging Large Language Models (LLMs) to power our own SOC infrastructure and offensive operations for years. We didn't just experiment with these models; we built full solutions around them to speed up the testing of large context datasets. We have effectively used human-in-the-loop guidance with earlier models to achieve results that the market is only now waking up to.
The difference is not the invention of the concept, but the engineering maturity of the agent’s autonomy. The emergence of high-context models like Mythos changes the dynamic for the general market, but for teams that have invested in proper code and infrastructure, this is an evolution, not a revolution. The implications are twofold: it democratizes high-level attack capabilities for low-skilled actors, and for experts, it provides a way to process millions of tokens to find attack patterns humans would miss. If your organization is not integrating these agentic capabilities into your testing cycles, you are failing to emulate the threat landscape of today.
From Script Kiddie to Prompt Kiddie
The most immediate danger of advanced agentic models is the lowering of the barrier to entry. In the traditional threat model, there was a distinct technical gap between a sophisticated red teamer and a "script kiddie." The script kiddie relied on copying Exploit-DB code, often failing to compile it or correctly configure the parameters. They posed a threat primarily to unpatched, legacy systems.
High-context AI bridges this gap. A novice with basic operational concepts can now instruct an agent to perform tasks that previously required deep understanding of network protocols or coding syntax. Mythos, specifically, changes this equation by requiring less guidance and less input from the human to gain access to a system.
Consider a standard scenario: identifying and exploiting a deserialization flaw. Traditionally, this required the attacker to understand the serialization format, find a gadget chain, and craft a malicious payload. With an agentic model like Mythos, the interaction changes dramatically:
- Traditional: The attacker reads documentation, writes Python code, debugs errors, and tests against the target.
- Agentic AI: The attacker uploads the HTTP request log and prompts: "Analyze this traffic, identify potential deserialization points, and generate a ysoserial payload for the CommonsCollections chain targeting this specific Java version."
The model interprets the request, analyzes the captured headers (e.g., Content-Type: application/x-java-serialized-object), identifies the environment, and writes the functional payload. The human operator moves from being a skilled craftsman to a supervisor. This "Prompt Kiddie" phenomenon effectively turns a low-skilled actor into a competent automated attacker. They may not understand why the payload works, but they possess the toolchain to make it work.
The Cognitive Advantage: Million-Token Context
Where agentic AI truly diverges from previous iterations—and from human capability—is in the management of context. Human penetration testers suffer from cognitive fatigue and information overload. We can only hold a finite number of variables in our working memory while trying to pivot through a compromised network.
Models capable of processing millions of tokens do not have this limitation. This allows for a depth of reconnaissance that is practically impossible for a human to replicate in a reasonable timeframe. We are observing these models making connections that are frequently missed during manual engagements because they can "see" the entire attack surface at once.
Pattern Recognition at Scale
During recent assessments, our team utilized high-context models to ingest massive datasets of internal JavaScript files and API documentation aggregated from a target's infrastructure. A human analyst might spot an exposed API key or a hardcoded credential after hours of grepping, but the model was able to correlate disparate pieces of data across thousands of files.
It identified specific functions handling user authentication communicating with legacy endpoints. By cross-referencing these endpoints against separate repositories of internal documentation, the agent inferred default administrative credential structures. The "attack path"—linking modern frontends to legacy backends—was established through the ability to read the entire estate simultaneously.
The Operational Reality: Precision Over Hallucination
A common misconception in the industry is that AI penetration testing is plagued by unmanageable false positives and hallucinations. This view often stems from relying on raw, unguided model outputs rather than engineered solutions. At Security Arsenal, we have proven that by using reasoning models coupled with robust control code, agents can self-test, confirm, write the Proof of Concept (PoC), and validate an issue autonomously.
When done effectively, a system driven by an experienced red teamer can outperform most traditional red teams. The human drives the direction and the tests, while the AI executes with accuracy. By forcing the model to double-check against facts, proof, and validation methods before reporting a finding, we effectively eliminate the hallucinations that critics cite. If the code is tested and effective, the results are accurate.
Securing Data: Scrubbing and Gating
Another frequent objection is the risk of data leakage when sending sensitive client data to AI models. This is a solved problem with the right architecture. Our design utilizes placeholders that automatically scrub sensitive data before it ever reaches a public model. Furthermore, the majority of our processing directs queries to local LLM models, negating the risk of exfiltration entirely.
We implement a strict gating system that tracks data flow, ensuring that no proprietary information leaves our controlled environment. This allows us to leverage the massive processing power of AI for offensive security without compromising client confidentiality.
The Evolution of Offensive Tooling
The impact of AI is not limited to autonomous agents; it is revitalizing the entire offensive toolchain. As security vendors integrate AI into their products, we are seeing older tools updated to a new level of sophistication. For example, the new Metasploit 6.5 includes amazing features for evasion and payload generation that were previously manual, time-consuming tasks.
However, off-the-shelf tools are only part of the equation. We have developed our own proprietary tools that leverage AI to write custom code capable of evading industry-standard defenses. These tools, which operate in memory and are never detected by EDR or Defender, represent the cutting edge of AI-assisted offensive development. While we will not disclose our trade-secret methods for hiding in memory on any device, the fact remains: AI allows us to generate unique, polymorphic code that bypasses signature-based detection and heuristic analysis.
Bottom Line
The hype around Mythos and similar models is just the visible tip of a much deeper shift. The commoditization of attack expertise is real, but the true power lies in combining AI with expert engineering to create precise, autonomous offensive systems that do not hallucinate and do not leak data.
What To Do Next:
- Adopt Engineering-First AI: Stop treating AI as a chatbot. Invest in reasoning models and build control code that enforces self-validation and proof-of-concept generation to eliminate false positives.
- Implement Data Gating: If you use public models, build an automated scrubbing layer that replaces sensitive data with placeholders before transmission. Prioritize local LLM deployments for sensitive operations.
- Leverage AI for Evasion: Move beyond basic exploits. Use AI to generate custom, in-memory payloads and evasion techniques that bypass standard EDR and Defender signatures.
- Update Your Toolset: Integrate AI-enhanced versions of standard tools like Metasploit 6.5 into your workflows to take advantage of advanced evasion and automation features.
- Pilot a High-Context Purple Team: Test your defenses against an AI-driven red team that utilizes million-token context to find obscure attack paths across your entire environment.
Related Resources
Security Arsenal Managed SOC Services AlertMonitor Platform Book a SOC Assessment Threat Intelligence Blog
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.