SEC595: Applied Data Science and AI/Machine Learning for Cybersecurity Professionals


Experience SANS training through course previews.
Learn MoreLet us help.
Contact usBecome a member for instant access to our free resources.
Sign UpWe're here to help.
Contact Us
Secure Your Fortress 2026 brought together practitioners, researchers, and educators for a day of honest, no-fluff conversation about where cybersecurity actually stands right now. The throughline that emerged across virtually every session was acceleration: attackers are moving faster, AI is reshaping both sides of the equation, and security teams are being rushed to decide how to keep up. Talks ranged from rethinking how we build detection programs to the very real risks of agentic AI and MCP-based tool ecosystems, revealing the cybersecurity industry at an inflection point.
The day's conversations kept returning to the same uncomfortable truth: the gap between how fast threats are moving and how slowly most programs are adapting is more about judgment than technology. The biggest risk facing security programs in 2026 is the accumulation of unexamined assumptions: what your detections actually cover, whether your timestamps can be trusted, and what your AI agents are really doing. Successful practitioners have stopped inheriting assumptions and started testing them.
The day opened with Rich Greene challenging the way the field teaches and talks about itself. Speaking about the sprawling list of terms and concepts that greets anyone new to cybersecurity, he said, "It's not a vocabulary list. It's a wall," and he pointed out the unspoken message that wall sends. Real security fluency, Greene argued, isn't about accumulating terminology. Fluency is the ability to see how the pieces connect: how a detection gap becomes a business risk, how a credential attack becomes a supply chain problem, how an unresolved team dynamic still shapes how threats get communicated two years later. Crucially, the practitioners who achieve that fluency don’t do it alone: "The job is to point at the board for the next person, so they don't have to take the long way around."
You can't protect what you can't see, and most teams have bigger blind spots than they realize. Nick Mitropoulos made the case that having a log source present doesn't mean you have the coverage you think. The right event IDs need to be enabled, the right fields need to be populated, and the data needs to actually reflect what's happening in your environment. Meanwhile, Eric Conrad pushed back on the prevailing assumption that encrypted traffic makes network-based detection irrelevant. Encryption hides payload content, but it doesn't hide behavioral metadata. Structural deviations in TLS sessions are detectable without decrypting a single packet, and they're characteristic of how major C2 frameworks operate. Endpoint detection and network visibility complement each other, and organizations that have abandoned one of the two are operating with a blind spot.
Some of the most impactful incidents come down to the basics going wrong. Bryan Simon put it plainly: "Time confusion is silent, costly, and potentially deadly when it comes to an investigation." His deeper point was that "time was never designed for accuracy," and security teams tend to inherit imprecision without questioning it. A 60-second skew between two log sources can invert the order of events in an attack chain, making lateral movement appear to precede the initial access that enabled it. Computer clocks can drift, daylight saving transitions can create log gaps or duplicate event windows, and cloud environments compound the problem across regions with different time zone configurations. The practical defense is straightforward: normalize all log timestamps to UTC before analysis begins, then verify NTP synchronization as part of your detection infrastructure baseline, and treat timestamp anomalies as a signal worth investigating. Most organizations haven't checked those assumptions recently.
At SYF 2026, John Hubbard observed something hard to miss: "The agents have arrived. They are here." Nearly every vendor, banner, and conversation on the show floor led with the topic of agentic capabilities. For defenders, the shift from single-turn AI interactions to agentic workflows, where you set an objective and the agent plots a path to it, is already part of production environments, touching every stage of the SOC from collection through incident response. Hubbard described these tools as "as close as anything to magic that we have right now." Security teams can now turn "I wish I had a tool for this" into a working solution in minutes, directly closing the gap in operations tempo with attackers. Dave Shackleford made a parallel case for AI in vulnerability management: with over 30,000 CVEs published annually and NIST announcing it can no longer enrich the full database, the burden of triage is shifting to organizations themselves. AI can reason over asset criticality, exposure data, and exploit telemetry at a scale no human team can match, freeing analysts to spend more time remediating and less time prioritizing.
However, it’s important to understand what you're building with and what its privileges are. Mark Baggett demonstrated this live. An AI agent was given strict instructions never to reveal credit card numbers, but then was socially engineered through a math tutoring prompt that required a 16-digit number from its records; the agent resolved the conflict between its two competing rules by running an UPDATE query that replaced every credit card number in the database with a random string. "In order to protect the credit cards, we have to destroy all the credit cards, and AI just did it," Baggett said. Principle-based constraints aren't sufficient for agent security. Practitioners building production workflows without understanding context management, temperature, or prompt sensitivity are accumulating risk they can't see until something breaks or is deleted.
The same capabilities that make AI agents powerful for defenders make them dangerous as an attack surface. An agent that can autonomously triage alerts, isolate machines, and remediate vulnerabilities is also an agent that holds credentials, calls APIs, and takes action at machine speed, with or without human oversight. Your security program must be ready to govern what you've already deployed.
The most urgent emerging theme of the day: agentic AI and MCP-based ecosystems are introducing a new class of identity and attack surface that most organizations aren't ready for. Ismael Valenzuela framed the core problem: agents "use the same tokens that humans use. They call the same APIs that humans may call, and they run on machine speed." In the average enterprise, non-human identities outnumber human identities roughly 50 to 1, and 80% of organizations report these agents acting outside their expected behavior (Cloud Security Alliance). The detections built for humans and service accounts don't translate to agents. Valenzuela's prescription was direct: "Treat agents as operators" with identity, lifecycle management, behavioral baselining, and accountability. The missing dimension in existing Zero Trust frameworks is intent: what is this identity trying to do, and does that match what it's supposed to do?
Seth Misenar specified the attack surface that makes this urgent. "The injection is in tool descriptions, in retrieved documents, in MCP server responses, in agent memory files … It's not merely in the prompt anymore." MCP servers expose their capabilities through tool descriptions written in natural language, making those descriptions themselves the primary attack surface. Approximately 43% of analyzed MCP servers have command injection flaws, 79% pass API keys through environment variables, and nearly 500 publicly exposed MCP servers have been found with no authentication whatsoever (BitSight). Adoption is outpacing security, and the industry may spend years catching up, but we know the pattern and the controls exist.
Modern investigations demand analytical discipline across open sources, digital evidence, and an information environment that is increasingly manipulated.
Dario Beniamini made the case that OSINT and digital forensics are more intertwined than how most practitioners treat them, and that the flow between them is genuinely bidirectional. In one concrete example, a Windows forensic image yielded Wi-Fi SSIDs and Gateway MAC addresses from the registry, which were then cross-referenced with Windows event logs for connection timestamps, which in turn were run through Wigle to geolocate where a subject was and for how long. In another example, mobile forensic databases that did not capture full Telegram group history were supplemented by bulk-acquisition of that data via open-source tools; investigators then pivoted back through OSINT to identify participants. The through-line across all three examples was the fact that entities extracted from a forensic container (usernames, emails, crypto wallets) become the starting point for OSINT enrichment, and that enriched intelligence feeds back into the forensic workspace for link analysis.
Chris Pizor brought a complementary warning: the line between information security and information integrity is blurring. This can mean that defenders working with threat intelligence or incident response communications are already encountering manipulated information as part of their operational environment, possibly without applying the same analytical standards to what they read as to what they investigate. AI has lowered the production cost of synthetic content to near zero, changing the scale at which disinformation operations can run. The analytical methodology for evaluating credibility hasn't changed — source, motivation, corroboration, distribution pattern, and emotional framing — but it is now more urgent to apply these criteria. As Pizor put it, modern influence operations are about steering.
Across a full day, the judgment problem showed up in every session. Teams need to confront whether they have actually tested what their detections cover, or just assumed they do. Leadership must know whether agents have been inventoried and governed, or just deployed. Investigators should read threat intelligence with the same rigor as they read forensic evidence, or else let manipulated information shape their conclusions unchallenged.
The fundamentals are still where most incidents live. Token theft doesn't require a zero-day. Skewed timestamps don't require action from an attacker. Detection gaps can become costly even when adversaries are unsophisticated. The sessions that drew the most engagement were the ones that highlighted assumptions defenders may have stopped questioning.
Real fluency is collaboration and mindset. The practitioners who are thriving understand why each piece of their program works, where it doesn't, and what the next conversation needs to be.


Launched in 1989 as a cooperative for information security thought leadership, it is SANS’ ongoing mission to empower cybersecurity professionals with the practical skills and knowledge they need to make our world a safer place.
Read more about SANS Institute