How Setu makes you Mythos-ready

The CSA's expedited briefing hands CISOs eleven priority actions for the AI vulnerability storm. Here is where a security graph carries the load, where it helps, and where the work is honestly yours.

In April 2026, Anthropic previewed Claude Mythos: autonomous discovery of thousands of zero-days across every major operating system and browser, working exploits without scaffolding, a 72% exploit success rate. Within weeks the Cloud Security Alliance, SANS, and the OWASP GenAI Security Project published an expedited strategy briefing, with eighty CISOs contributing and another two hundred fifty reviewing. Its core observation is hard to argue with: mean time from disclosure to exploitation has collapsed from 2.3 years in 2018 to roughly nine hours in 2026, and the CVE-and-patch pipeline was never designed for that world.

The briefing distills the response into eleven priority actions, from pointing coding agents at your own code to standing up a permanent VulnOps function. Boards are reading it. And vendors are already publishing coverage tables claiming eight, nine, ten of the eleven. (We wrote about the post-Mythos boardroom moment in May, in After Mythos: what an honest defender's stack looks like this fall. This piece is the companion: the CSA's action list, row by row.)

We are going to do something different: map Setu against each action honestly. Several of the eleven are organizational, and no product covers them. A few belong to your appsec program, not your exposure program. The rest are graph problems, and that is where we live.

The mapping

Six of eleven carried or led by the graph, three shared, two honestly yours. Here is the reasoning behind the rows that matter most.

The inventory is the graph

The briefing's triage questionnaire asks a question most programs fail: "Are your crown jewels explicitly tracked and current? Not theoretically important systems. The actual few that matter most, and their main dependencies."

That question is unanswerable in a spreadsheet and native in a graph. Setu's entity graph unifies assets, identities, findings, and controls from the tools you already run, and crown jewels are not a tag someone remembered to apply. They are nodes whose dependencies and reachability are computed, continuously, from the same telemetry that feeds detection.

The same structure fixes risk reporting, which the CSA rates a critical gap. Static severity scores measure generic exploitability; the briefing asks for risk models that measure exploitability in your environment, against your controls. That is the difference between a CVSS 9.8 on an isolated dev box and a CVSS 7.0 sitting on the path between an internet-facing service and your payment database. Score by graph position and the false criticals fall away. Report risk as open versus closed, not patched versus unpatched, and the board conversation changes shape.

Blast radius is a graph property

The briefing is blunt about detection in a Mythos world: signature-based detection fails by definition against a vulnerability that did not exist an hour ago. What remains is architecture. Segmentation, egress filtering, least privilege: every boundary increases attacker cost.

Setu does not segment your network. What the graph contributes is the answer to "which boundary, first?" Attack-path analysis ranks the junctions that carry the most paths to the things you cannot afford to lose. That is Pagerank for Posture: hardening effort ordered by structural payoff rather than by whichever finding shouted loudest this week.

Deception: we tell you where, you decide what

Priority action nine gives deception its overdue promotion, and the reasoning is sound: decoys are attack-tool independent. A signature does not exist for a zero-day, but nobody has a legitimate reason to touch a decoy, whatever exploit got them there. Against agentic attackers that probe every path exhaustively, decoys also make every probe expensive.

The unsolved problem in deception has never been building decoys. It is knowing where to put them. Placement decided by intuition produces decoys that nothing ever touches. The graph already knows the choke points on real attack paths, the identities worth imitating, the shares an intruder would enumerate first. Setu's Placements recommend where deception earns its keep and why, and you deploy with whatever decoy or honeytoken tooling you prefer.

Machine-speed response needs authority, not just speed

The briefing names the uncomfortable constraint: "containment authorization latency was designed for human-paced threats." The bottleneck is not detection speed. It is the meeting where someone decides whether the SOC is allowed to isolate the host.

The CSA's prescription is pre-authorized containment actions and playbooks that execute at machine speed. That is precisely what Prescriptions are built around: per-action authority stages agreed in advance, approval gates where a human genuinely must decide, rollback for every action so containment is never a one-way door, and execution through the enforcement points you already own. Speed without authority is a demo. Authority without rollback is a liability. You need all three before the first machine-speed incident, not after.

VulnOps: the half that will actually drown you

The final priority action calls for a permanent vulnerability operations function, staffed and automated like DevOps. Read the fine print and it splits in two. Discovery, finding the flaws, is being solved at industrial scale by the model labs and your own agents. Triage is the half that drowns you: the briefing warns of hundreds of critical disclosures per month against a CVE system that "may not scale," and threat intelligence that structurally lags discovery, because an AI-found vulnerability has no KEV entry by definition.

You cannot feed-rank your way out of that. Setu's answer is validation over trust: dedupe the flood across sources, test which findings are actually exploitable in your environment, dispatch the survivors to owners with the context to fix them, and keep signed evidence packs so the audit trail exists the day a regulator or board asks. The briefing frames the coming EU AI Act era as a shift in the standard of reasonable care. Proof of diligence is about to matter as much as diligence.

What we will not claim

Pointing agents at your own code is an appsec practice, and mandating agent adoption is a leadership decision. No exposure platform does either for you, whatever the coverage table says. And where our coverage is partial, the table says so: we would rather you know exactly which half of an action you still own than discover it mid-incident.

The CSA closes with a line worth keeping: being Mythos-ready "is not about reacting to one model or announcement. It is about permanently closing the gap between how fast vulnerabilities are found and how fast your organization can respond." That gap is a graph problem. Bring us your program and the eleven actions, and we will map them against your estate, honestly, including the rows where the answer is "not us."