Eric Fourrier is the CEO of GitGuardian, a secrets visibility and intelligence platform for enterprises.gettyAnthropic’s Claude Mythos Preview has shown two concerning things at once: A frontier model can discover large volumes of previously unknown vulnerabilities, and compress the time required to develop exploits for newly disclosed bugs from days or weeks to hours.The security community moved fast to respond. Within days, the Cloud Security Alliance, SANS and the OWASP GenAI Security Project published a briefing, The AI Vulnerability Storm, on how to run a security program when exploit discovery happens at machine speed. It is a serious, well-sourced document, and one thread running through its priority actions is worth focusing on.Several of its highest-priority controls rest on the same premise: Some attackers will get in, so make that access costly. Phishing-resistant authentication raises the bar on stolen credentials, and the briefing puts honeytoken-based detection on a 90-day clock specifically to catch intruders already inside. A leaked API key, service account token, OAuth grant or standing cloud permissions turns one foothold into lateral movement, data theft and business impact. The exploit may get an attacker through the door, but credentials decide how many rooms are unlocked. That is also the part security teams can still control: what a stolen credential is worth after an attacker has it.The 2026 Verizon DBIR makes that distinction clear. Exploiting software vulnerabilities is now the leading way breaches start, accounting for 31% of initial access. But credentials still show up throughout the breach path: Credential abuse appears in 39% of breaches, stolen credentials remain a major action category, and compromised credentials continue to drive a large share of basic web application attacks. This is a stark reminder that entry is not impact. When entry is cheap, leaked credentials shorten the attack path. In the 2025 Salesloft Drift–Salesforce campaign, attackers used stolen OAuth tokens from Salesloft’s Drift integration to reach Salesforce instances at more than 700 organizations, then mined exported data for embedded secrets that could unlock further systems. The pattern repeated in June 2026, when attackers breached Klue through a long-disused credential, harvested customer OAuth tokens connected to Salesforce, and pulled data from nearly 200 organizations.AI Multiplies The Credentials You Have To GovernOrganizations use the credential layer to limit blast radius, but AI is making that layer larger, faster-moving and harder to govern. Every useful AI workflow connects to something: SaaS apps, data, tools and, increasingly, other agents. Each connection depends on a credential, whether that means an API key, OAuth grant, service account or token. As agents proliferate, these machine-to-machine trust relationships can quickly outpace the teams responsible for discovering, scoping, owning and rotating them.Those credentials do not stay in source code either. The 2026 GitGuardian State of Secrets Sprawl report found that about 28% of secrets incidents originate exclusively outside code repositories, in collaboration and productivity tools such as Slack, Jira and Confluence, and that more than half of those incidents are critical. That is probably a floor; AI-assisted workflows push secrets into logs, prompts, traces, plug-in configurations, tool outputs and other runtime artifacts. The briefing’s “10 Questions” diagnostic asks whether organizations govern the agentic supply chain, including MCP servers, plug-ins and skills. The next question is which credentials those supply chains hold, where they live and who owns rotation.The scale problem is already here; CyberArk’s State of Machine Identity Security research found that 50% of organizations have reported breaches tied to compromised machine identities, including API keys and TLS certificates. Many security teams still lack a reliable inventory of the nonhuman identities they already have, let alone the new trust relationships that agentic workflows are beginning to introduce.​Put Assume-Breach Controls On The 45-Day Track​The briefing lays out a phased plan with deliberately aggressive timelines: patch readiness and risk-model updates starting this week, asset inventory and environment hardening this month, and deception and automated response inside 90 days. Credential containment belongs on the same track because entry and impact are separate problems. Patch readiness helps control entry, but long-lived tokens, overprivileged service accounts and unmanaged OAuth grants still determine how far stolen access can go.A 45-day credential track should do four things: identify abusable identities, cut off stolen access fast, catch attackers using credentials you did not know were exposed, and keep remediating exposed secrets before they accumulate.1. Inventory your nonhuman identitiesYou cannot contain what you cannot see. Priority Action 7 sets a 90-day target for full asset inventory; start with the most privileged service accounts and tokens, where one exposed credential can do the most damage. Extend that inventory beyond repos and vaults to developer machines, where .env files, cloud credential stores and AI-agent configs increasingly expose more credentials. Automation and AI-assisted discovery can help compress the timeline.2. Measure time-to-revokeWhen a credential is confirmed as exposed, track how long it takes to fully invalidate that access. The secret must be rotated, dependent systems updated and any post-exposure access reviewed. Start with privileged accounts, production secrets and third-party access keys.3. Deploy honeytokensThe briefing makes this Priority Action 9 on a 90-day clock. Pull it forward. Canary credentials detect intruders using stolen access regardless of how they got in, which is exactly the signal this posture needs.4. Remediate exposed secrets continuouslyImprove hygiene before the breach. Keep rotating and removing critical credentials from the places attackers harvest them, including developer laptops and code repositories. When a honeytoken fires, that same discipline is what lets you rotate fast.The organizations that get this right make stolen access less useful. That is the point of assume breach. You cannot control every exploit path, but you can control what is stolen.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?