The Mythos-ready briefing landed last week, co-signed by Jen Easterly, Bruce Schneier, Heather Adkins, Rob Joyce, Chris Inglis, Phil Venables, and 60+ other CISOs from Google, Snowflake, Atlassian, and organizations like the NFL and TransUnion. Among the controls they named as critical for the AI vulnerability era were secrets rotation, non-human identity governance, early detection of compromise, and honeytoken-based deception. If you've been pushing for more budget for better secrets security, this is the document to put in front of your CISO.
What the paper says about credentials
The briefing is a response to Anthropic's Claude Mythos Preview announcement, which reported autonomous discovery of thousands of zero-days across every major operating system and browser with a 72% exploit success rate. The paper lays out 11 priority actions, a risk register, and a 90-day plan for CISOs. Credentials underpin nearly every control it calls out.
In the Key Takeaways, the authors name secrets rotation alongside segmentation, egress filtering, Zero Trust, and phishing-resistant MFA as mitigating controls that limit blast radius when exploitation occurs. The risk register tags "Unmanaged AI Agent Attack Surface" as CRITICAL, pointing to privileged agents operating outside existing control frameworks.
Priority Action 8 ("Harden Your Environment") mandates phishing-resistant MFA for all privileged accounts and locking down the dependency chain. Priority Action 9 ("Build a Deception Capability") calls for deploying canaries and honeytokens, layered with behavioral monitoring and pre-authorized containment. The executive briefing section frames early detection of compromise as a metric boards should be tracking. This briefing is a rare alignment of industry leadership putting credential security squarely on the critical-controls list.
Why a Mythos world makes credentials matter more
There's a common misreading of the AI vulnerability story, which is that zero-days become the dominant threat and everything else fades. The paper's own Appendix A pushes back on that. The authors note that the historical collapse in time-to-exploit has not produced a proportional rise in exploitation impact, and that most consequential recent breaches came from credential abuse, social engineering, or supply chain compromise rather than zero-day exploitation.
The 2025 Verizon DBIR backs this up. Stolen credentials remain the leading initial access vector at 22% of all breaches, and 88% for basic web application attacks. Machine identities are now involved in some stage of 68% of IT security incidents.
Layer Mythos-class capability on top of that, and valid credentials become the fastest way in. When zero-days are cheap, they accelerate lateral movement after initial access. They don't replace credentials as the entry point. That's how the Snowflake breach in 2024 hit 165 organizations from credentials that had been sitting in infostealer logs, some since 2020. MFA wasn't enforced, rotation hadn't happened, and old credentials were still valid — no novel exploit needed.
AI is accelerating the credential sprawl that underlies all of this
That risk is accelerating. AI drives credential exposure on two fronts, volume and surface area. As the paper notes, higher code output with less consistent review increases the number of vulnerabilities that ship. The same velocity drives a parallel explosion in credential creation.
Our State of Secrets Sprawl 2026 report found 29 million new hardcoded secrets exposed on public GitHub in 2025, a 34% year-over-year increase and the largest single-year jump on record. Credentials tied specifically to AI services surged 81% year-over-year.
And 28% of secrets-related incidents in our 2026 data originated entirely outside source code. They showed up in CI/CD systems like GitHub Actions and GitLab runners, in collaboration surfaces like Slack, Jira, and Confluence, and on developer machines.
Those are now the same surfaces AI agents read, summarize, and act on as part of day-to-day workflows. The paper's "10 Questions" diagnostic asks whether organizations have disciplined control of their agentic supply chain, including MCP servers, plugins, and skills. The credential question sits directly underneath: what secrets do those systems hold, where do they live, who owns them, and how fast can they be rotated when something goes wrong?
In most enterprise environments, non-human identities already outnumber human users by a ratio of roughly 25-50x. Very few organizations have an inventory of the ones they already have, let alone the ones AI agents are creating at scale.
What security teams actually need
Security teams need visibility everywhere credentials actually sprawl: repos, CI logs, container layers, tickets, chat threads. That's a solvable problem. The harder part is connecting each exposed secret to the non-human identity behind it and figuring out which services, workloads, or automations depend on it. Without that context, triage stalls, and an exposed credential gets used before anyone can act on it.
Ownership is where most of this work breaks down. When a credential is exposed, the question "who owns this?" usually doesn't have a clean answer. The developer who committed it may have left the team. Often, the service it authenticates runs in a different group's infrastructure entirely. The rotation path may cross three systems that were never designed to coordinate with each other. In practice, that means the incident sits in a queue while three teams figure out whether it's theirs. Every hour in that queue is an hour the credential is live and usable. That's the exposure window.
Non-human identities compound the problem. A service account created for a CI pipeline two years ago may have no human owner on record. No one's inbox to land in, no runbook to follow.
Most security programs already struggle to detect exposed credentials. They don't even touch ownership and response, which is the gap GitGuardian was built to close. GitGuardian gives teams continuous secrets detection across source code and other places where secrets appear. That includes CI/CD systems like GitHub Actions and GitLab task runners, collaboration platforms like Slack and Jira, and developer environments down to the laptop. It surfaces exposed credentials where modern work actually happens, not just where security teams wish it did. From there, NHI discovery and ownership mapping connect exposed secrets to the service accounts, API keys, and machine identities that power agentic systems and automation at scale.
A case for moving credential hygiene up the priority list
Containment is the whole game once time-to-exploit collapses to hours. You can't afford to find credential exposure days or weeks after the fact. A secret sitting in Slack or a build log doesn't show up in a vulnerability scan. An API key tied to an agent workflow still expands the attack surface. A service credential without an owner still slows every remediation step that follows.
The paper draws a clear line through its 11 priority actions. With exploitation becoming both faster and more automated, response speed and blast-radius reduction move to the center. Secrets rotation, non-human identity governance, phishing-resistant MFA, and honeytoken-based detection belong at the front of the list as core resilience controls. They shape how quickly an organization can contain misuse once an attacker gets in, or once an agentic workflow is abused.
Given what the data shows, those controls deserve to be on the 45-day track alongside environment hardening, not grouped underneath it. In our longitudinal dataset, 64% of secrets leaked in 2022 still hadn't been revoked as of 2026. The paper warns that time-to-exploit has collapsed to hours. Those two numbers don't coexist safely in the same environment.
GitGuardian directly supports that shift. Secrets detection helps teams find exposed credentials before attackers do. Rotation signals and remediation workflows push incidents toward closure instead of letting them linger.
NHI discovery and control help organizations understand which machine identities exist, what they can access, and who's responsible for them. GitGuardian Honeytokens add an early warning layer that surfaces credential misuse before a broader incident unfolds. That maps directly to Priority Action 9 in the paper, which calls for honeytoken deployment, behavioral monitoring, and pre-authorized containment. The goal is a response that executes at machine speed.
If you're building your 90-day plan from the Mythos briefing, credential security deserves to move up the list. Hardening, detection, and response all come down to the same question: when something moves, how fast can you contain it? The organizations that come through this well will be the ones that had that answer before they needed it. Our 2026 State of Secrets Sprawl report has the full picture.
