Mythos-Ready Security Operations

What Mythos is, and what it isn't.

On April 7, 2026, Anthropic announced Claude Mythos Preview, a model so capable at autonomous vulnerability discovery that the company chose not to release it to the public. Instead, it launched Project Glasswing — a coordinated program to put the model in the hands of roughly 40 defensive partners first.

In Anthropic's own testing, Mythos identified and exploited zero-day vulnerabilities in every major operating system and every major web browser when directed to. Many of the bugs it surfaced were 10 or 20 years old. The oldest was a 27-year-old vulnerability in OpenBSD — an operating system whose entire reputation rests on its security engineering. On a single benchmark (Firefox), the prior best model produced 2 working exploits. Mythos produced 181.

Six days after Anthropic's announcement, the UK AI Security Institute (AISI) published independent evaluation results. On a 32-step attack range called "The Last Ones" — covering reconnaissance through full network takeover, estimated to require 20 hours of focused work from a human professional — Mythos became the first model in history to complete the full chain. It did so in 3 of 10 attempts. The previous best model averaged 16 of the 32 steps and never reached the end.

Two distinctions matter here, and the security community is mostly aligned on both.

First, Mythos is not the threat — it is the announcement of a threat class. Rob T. Lee at SANS put it this way: AI-driven vulnerability discovery has been accelerating for more than a year. Mythos compresses the timeline further, but the capability is not new, and waiting for the next major announcement is not a strategy. Open-weight models without safety guardrails will close the gap, and the cost and skill floor for autonomous vulnerability discovery has permanently dropped.

Second, defense in depth still works. Rich Mogull at the CSA noted that in Anthropic's own testing, Mythos found exploitable Linux kernel vulnerabilities — and after several thousand scans, could not remotely exploit a single one of them. The hardening that defenders have done over years held. The cost of turning a bug into a working attack is still a function of how much defensive engineering sits between the bug and the asset. The fundamentals matter more than ever, not less.

"Even with active human defenders and alerts, the speed asymmetry is what should keep you up. The window between vulnerability discovery and weaponization has collapsed into hours." — Rob T. Lee, Chief AI Officer, SANS Institute · via Cyber Magazine

Why this lands hardest in the SOC.

The AI Vulnerability Storm is not a single-domain problem. It increases load on vulnerability management, AppSec, identity, and incident response. But the CSA/SANS/OWASP briefing is explicit about where the cost of unpreparedness lands hardest. It is the SOC.

The reasoning is simple. Every other domain produces alerts. When AppSec finds exposed weaknesses faster, the SOC sees more signal. When vulnerability management surfaces a larger backlog, the SOC sees the exploitation attempts. When identity sees more attempted abuse, the SOC catches the lateral movement. All roads from the AI Vulnerability Storm pass through security operations.

The SOC team that was already running at capacity does not get a longer runway under these conditions. It gets a faster adversary and more signal to process. The window between initial access and meaningful damage collapses. Mandiant's M-Trends 2026 report measured attacker handoffs inside compromised networks at 22 seconds. In 2022, that window was over eight hours. The mean time to exploit vulnerabilities has dropped to an estimated negative seven days, meaning exploitation is now routinely occurring before a patch even exists.

Neither capacity nor speed is solvable by hiring. There are an estimated 4 million unfilled security roles globally; even if every one were filled tomorrow, the math of human-paced investigation versus machine-paced attack does not improve. Both problems are structural. Both require structural answers.

"The challenge facing cybersecurity isn't about awareness or capability, it's about physics. A sophisticated attack group can compromise a network, steal data, and disappear in under an hour. Even the best security teams are celebrating reducing mean time to response from days to hours. The math simply doesn't work. You can't defend against machine-speed attacks with human-speed responses, no matter how skilled your team is." — Yonatan Striem-Amit, CTO and Co-Founder, 7AI · via 7AI Blog

That is why the consensus among the people writing about Mythos — from CSA to SANS to AISI to CISA — is converging on the same operating model: AI agents under expert human oversight, running at machine speed, with the SOC restructured off human schedules. That is what the rest of this page is about.

What operationally changes in a Mythos-ready SOC.

The CSA briefing is the doctrine. Translating it into operational reality means restructuring the SOC around four practical shifts. None of them are theoretical. All of them are achievable inside 90 days for organizations that act now.

1. Every alert gets an end-to-end investigation.

In a traditional SOC, alerts queue. In a Mythos-ready SOC, AI agents pick up alerts the moment they fire, enrich them with context from across SIEM, EDR, identity, cloud, email, and DLP, and produce a structured investigation with evidence and reasoning. The analyst reviews, escalates, or closes. Nothing sits waiting for someone to have time.

This is the capability AI SOC platforms have been building toward for two years. It is now table stakes. The 7AI Agentic Security Platform has processed more than 7 million alerts this way and saved customers over a million analyst hours — the equivalent of nearly 521 full-time analyst-years. The math of human-paced triage no longer holds, and the consensus document is direct that this shift is non-optional.

2. Threat hunting runs continuously, not on a calendar.

The CSA briefing identifies behavioral hunting as the critical detection layer when AI-discovered zero-days do not trigger signature-based controls. Quarterly hunts were built for a world where attackers moved on monthly cycles. That world is gone.

In a Mythos-ready SOC, AI agents execute hypothesis-driven hunts across 90+ days of telemetry in minutes. New hunts go up the moment a CISA advisory drops or a fresh CVE gets disclosed. Last week's hunches run every night. This is covered in more depth in section 05 below.

3. The SOC operating tempo moves off human schedules.

The lagging indicator most security programs report on is mean time to respond. That number matters, but the leading indicator the CSA briefing implicitly endorses is different: whether the program has moved its core operations off human schedules. Investigations run continuously, hunts run continuously, detection engineering runs continuously, response runs continuously. Analysts move from triage to hunting, adversary modeling, board reporting, and the strategic work that requires judgment.

4. AI operates under expert human direction, not autonomously.

CISA's May 2026 joint guidance on agentic AI, published with international partners, is explicit on this point: autonomy equals risk. AI agents in security operations must operate with least-privilege access, behavioral monitoring, and clear governance. The CSA briefing concurs: human-centered oversight for agentic AI decisions is essential, especially in vulnerability triage and incident response.

This is why operating models matter as much as platforms. At 7AI, the PLAID operating model — People-Led, AI-Driven — is built around dedicated AI Security Engineers: senior 7AI security practitioners assigned to each customer who configure the platform to the environment, tune the agents to local detections and policies, and stay accountable for outcomes. PLAID Elite adds 24x7 managed coverage by expert analysts. The principle is consistent with the CSA and CISA guidance: AI executes; people lead.

"If a security operations center is still fundamentally dependent on human analysts manually triaging alerts, it is not equipped for what is coming. Not because those analysts aren't skilled. They are. But because the math doesn't work. The only logical response is to match machine-speed offense with machine-speed defense. AI agents that can ingest every signal, correlate across domains, initiate investigations immediately, and surface only what requires human judgment. Not a copilot that waits for someone to ask it a question. An agent that runs." — Lior Div, CEO and Co-Founder, 7AI · 7AI Blog, April 2026

How these shifts map to the CSA priorities

When the zero-day doesn't trigger, hunting is the only layer left.

There is one line in the CSA briefing that should stop every SOC leader cold:

"When a zero-day lands and EDR does not fire, the only remaining detection layer is behavioral hunting." — "The AI Vulnerability Storm: Building a Mythos-Ready Security Program" · CSA / SANS / OWASP

Read it twice. The briefing is saying that the era of trusting signature-based controls to catch the next wave of vulnerabilities is over. When AI can find and weaponize exploits faster than vendors can patch them, the security stack does not get to assume that the alert will fire. The detection layer that catches the attack is the one looking for behavior, not signature. And in most SOCs, that layer is run by humans on a quarterly cadence.

What traditional hunting assumes — and why those assumptions break.

The traditional threat hunting model assumes three things that no longer hold.

It assumes the hunter has time. A skilled analyst forms a hypothesis, queries the data, refines, pivots, and writes it up. The good ones do this for weeks at a stretch. Most SOCs do not have that time anymore. The alert backlog gets worse when Mythos-class capabilities go mainstream. The hunter who used to spend three weeks on a hypothesis now has three days, then three hours, then no time at all because Tier 1 is on fire.

It assumes the hunt is periodic. Quarterly hunts were built for a world where attackers moved on weekly or monthly cycles. Dormant known-but-unfixed issues can be weaponized in hours when AI tools are in play. A hunt that runs four times a year cannot keep up with a threat surface that moves four times a week.

It assumes the data is in one place. Real hunting reaches across SIEM, EDR, identity, cloud, email, DLP, and the application layer. The friction of pulling all of that into one query, manually, on every hunt, is what keeps hunting from happening at all.

What AI-driven hunting actually does differently.

A Mythos-ready hunting capability has to break all three assumptions. It has to run continuously, reach across the data wherever it lives, and not require a human to be available for every hypothesis.

• Hypotheses run continuously. AI agents take the kind of hypothesis a senior hunter would write on a whiteboard and execute it across 90+ days of telemetry in minutes. Last week's hunches run every night.
• The hunt reaches into the data, not the other way around. AI agents query SIEM, EDR, identity, cloud, email, and DLP directly. No manual data normalization, no waiting on data engineering, no copying queries across five consoles.
• Findings come with evidence and reasoning. When an AI agent surfaces something interesting, it does not just produce a hit. It produces the hypothesis, the data it ran against, the evidence it found, and the reasoning. Analysts review, escalate, or close — they do not start from scratch.
• Humans stay in the loop where it matters. Analysts decide what to hunt for. AI Security Engineers tune hypotheses and detection logic to the environment. The model is people-led by design.

The compliance angle: NIST SP 800-53 Rev. 5 RA-10.

Continuous threat hunting is not just a Mythos-era best practice — it is increasingly a compliance requirement. NIST SP 800-53 Rev. 5 control RA-10 requires federal agencies and many regulated industries to "establish and maintain a cyber threat hunting capability to search for indicators of compromise in organizational systems and detect, track, and disrupt threats that evade existing controls."

That language maps almost exactly onto what the CSA briefing now describes as the minimum Mythos-ready posture. For regulated industries — healthcare, financial services, federal contractors — the compliance driver and the operational driver have converged.

What this looks like in production.

The clearest way to show what Mythos-ready operations look like is to walk through a real campaign 7AI surfaced this spring — one that every layer of the standard enterprise security stack missed, and that the PLAID model caught.

"Right off the bat we've seen an 80% reduction in tier 1 analyst time. As those tickets progressed through the funnel, we've seen easily a 95-99% reduction in the tickets that human beings have to look at. Incredible outcome. We believe this is the world's largest, most successful AI SOC deployment." — Mike Baker, Vice President IT & CISO, DXC Technology · 7AI customer

What continuous hunting looks like in the platform.

The capability that ran the CRXfiltrate IOC sweep is the same one available to every 7AI customer — the threat hunting module that lets AI agents execute hypothesis-driven hunts across SIEM, EDR, identity, cloud, email, and DLP telemetry. Below: how it surfaces hunt suggestions, runs cross-system correlation, and produces evidenced findings analysts can act on.

7AI Platform · Threat Hunting Module

What it looks like when alerts move into the investigation layer.

The other half of the Mythos-ready posture is what happens to alerts the moment they fire. Swarming AI agents enrich, correlate, and reach end-to-end conclusions before anyone sits down at a keyboard. Analysts see the conclusion, the evidence, and the reasoning — and decide what comes next.

7AI Platform · Investigations

A regional health system meets RA-10 — and catches an active campaign in the first weeks.

A multi-hospital health system signed on with 7AI in part to satisfy NIST SP 800-53 Rev. 5 RA-10 and the HIPAA Security Rule requirement to protect against reasonably anticipated threats. The compliance driver was real. The operational outcome was more meaningful.

Inside the first weeks of the engagement, 7AI's continuous hunting surfaced a multi-wave phishing campaign actively targeting the customer's environment that had been missed by existing controls. A 7AI threat hunter, working alongside the customer's AI Security Engineer, identified the broader pattern and the customer's incident response team contained it. The customer was able to redirect three planned threat-hunter hires to other strategic work because the capability was now in place.

The numbers underneath.

The CRXfiltrate hunt and the health system catch are individual stories. They sit on top of a continuous production layer that is now running at meaningful scale.

Frequently asked questions.

What is a Mythos-ready security program?

It is the operating model recommended by the Cloud Security Alliance, SANS, and OWASP in their joint briefing "The AI Vulnerability Storm: Building a Mythos-Ready Security Program." It is a structural shift in how security operations run, designed to match the machine speed of AI-driven attacks. Core shifts include continuous exposure management, behavioral threat hunting, machine-speed alert investigation, and AI agents operating under human oversight.

Why does the SOC need to change for Mythos?

The CSA briefing identifies the SOC as where the cost of unpreparedness lands hardest. AI-driven offense collapses the time between vulnerability discovery and exploitation, which increases alert volume while shortening the window to make decisions. Human-paced SOC operations cannot absorb that shift by working harder. Capacity and speed are both structural problems.

Is Mythos actually a threat today, or just a research result?

It is both, but the threat is not really about Mythos itself. Anthropic has kept Mythos restricted under Project Glasswing. The threat is that AI-driven vulnerability discovery as a capability class is here, open-weight models without safety restrictions are emerging quickly, and the cost and skill floor for autonomous vulnerability discovery has permanently dropped. Mythos is the announcement of a threat class, not the threat itself.

How does behavioral hunting actually catch what signature-based detection misses?

Signatures detect what attackers have done before. When a zero-day lands and EDR does not fire, behavioral hunting looks for what attackers do once they are inside — credential abuse, lateral movement, unusual data access, privilege escalation patterns. These behaviors do not require a signature to detect. The CSA briefing calls behavioral hunting "the only remaining detection layer" in the Mythos era.

How does this relate to NIST SP 800-53 RA-10?

NIST SP 800-53 Rev. 5 control RA-10 requires organizations to establish and maintain a cyber threat hunting capability. The language maps almost exactly onto what the CSA briefing now describes as the minimum Mythos-ready posture. For regulated industries — healthcare, financial services, federal contractors — the compliance driver and the operational driver have converged.

What is PLAID?

PLAID is 7AI's operating model: People-Led, AI-Driven. Every customer gets dedicated AI Security Engineers who configure the platform to their environment, tune agents, and stay accountable for outcomes. PLAID Elite is the managed service tier that adds 24x7 overwatch by 7AI's expert analysts. The principle is consistent with CSA and CISA guidance: AI executes; people lead.

How quickly can a SOC adopt these shifts?

The CSA briefing's timeline runs from this week to 12 months, depending on which capability is being stood up. The SOC-side pieces — continuous behavioral hunting, end-to-end investigation of every alert, machine-speed response — are achievable on the briefing's timeline. 7AI deployments are typically in production within seven days, but the right pace is the one that fits your environment.