Claude Mythos Preview has shifted the cybersecurity question from what advanced AI might do to whether organizations can respond at the speed it enables. The model’s reported ability to find obscure flaws and chain attacks across networks is striking. The practical test now is whether security teams can identify exposed systems, rank the risks and reduce them before attackers act.
That challenge reaches beyond one model. AI-assisted attacks are already putting pressure on the time defenders have to investigate and remediate vulnerabilities. For organizations, the task is to build a security program that can absorb a flood of findings and act on the ones that matter quickly.
AI is compressing the time available to respond
Stanford’s 2026 AI Index reports that AI agents solve cybersecurity problems 93% of the time, compared with 15% in 2024. In a separate evaluation, the UK’s AI Security Institute found that Claude Mythos Preview could complete a simulated 32-step attack against a corporate network and retrieve data hidden on target machines. The scenario was estimated to take a human about 20 hours.
Those results point to a change in how attacks can be assembled. An AI system can connect reconnaissance, exploitation and privilege escalation into a sequence, with less need for a person to guide each step. Mythos also found and exploited vulnerabilities autonomously in trials, according to the source article. Tasks that once demanded days of specialist effort may be completed much faster.
That pace exposes a mismatch between attackers and many organizations. A Cloud Security Alliance report describes attackers working at machine speed while defenders remain constrained by organizational processes. When patching takes weeks or months, a vulnerability may be targeted well before a fix reaches every system.
Widely used operating systems, browsers and network devices are especially consequential because a flaw in them can affect many organizations. Email has become another pressure point: attackers are using AI-enabled prompt injection phishing attacks to deceive employees into sending money to fraudulent accounts. The range of targets makes it harder for security teams to treat AI-enabled attacks as a narrow concern.
More vulnerability reports demand better triage
The volume of reported vulnerabilities is rising alongside the capacity to exploit them. Google’s Threat Intelligence group reported that monthly vulnerability disclosures grew from about 5,000 in January to more than 10,000 in July 2026. Its figures also show the average number of vulnerabilities exploited in the wild increasing from roughly 10.5 per month in 2025 to about 18 per month in early 2026. Zero-day exploitation rose slightly as well.
When the number of findings grows, patching everything in order of discovery is a poor use of limited time. Teams need to prioritize based on evidence of exploitation, the importance of affected assets and the potential impact of compromise. Google and other researchers expect the trend to continue as AI helps discover and exploit flaws more quickly.
That shifts the center of gravity toward continuous exposure management. Organizations need current asset inventories, ongoing vulnerability scans and a way to connect technical findings with business risk. Tools from vendors such as Wiz, Tenable and Microsoft Defender for Cloud can support this work, but the process matters more than any one product: identify what is exposed, determine which weaknesses are most urgent, then route fixes to the people who can apply them.
Security Boulevard describes exposure management programs that map assets, vulnerabilities and misconfigurations, then prioritize and remediate the riskiest issues. AI can help sort findings and suggest actions, but teams still need to verify that a proposed fix works and that the affected asset has actually been addressed. The goal is a response pipeline that turns useful intelligence into changes across real systems.
Defenders can put AI to work
AI’s ability to accelerate security work also gives defenders a way to improve their response. Anthropic’s Project Glasswing is an early example. In early April 2026, roughly 50 organizations were testing Mythos Preview. Within weeks, participants had run it across their codebases and found more than 10,000 high- or critical-severity flaws, according to Anthropic’s account of the program. Anthropic said it would extend access to about 150 additional organizations in sectors including power, healthcare and communications.
The program is intended to help organizations develop AI-assisted security workflows. Participants share practices, coordinate vulnerability triage and use AI to propose fixes and strengthen code before flaws are exploited. This makes the initiative more than a test of a model’s offensive capabilities: it also offers a view of how AI could fit into everyday defensive work.
Possible uses include generating patches, converting older code into memory-safe languages and running AI red teams against an organization’s own systems. Anthropic says partners are using Mythos to generate fixes, while its Claude Security tool scans code and suggests patches with a variant of the model. Security teams can assess similar approaches alongside automated penetration testing, asset discovery and vulnerability management.
Automation can shorten the path from detection to remediation, but it doesn’t remove the need for human review. A patch that fixes one issue may introduce another, and a scan cannot confirm that every relevant system has been found. AI works best as part of a process with clear ownership, validation and follow-up.
Security basics still determine the damage
More capable tools don’t make foundational controls less useful. The UK’s AI Security Institute says Mythos’s capabilities underline the value of regular patching, strong access controls, secure configurations and comprehensive logging. Those practices limit the number of easy opportunities available to an attacker and help teams see what happens when defenses fail.
Shorter exploit windows may require more frequent updates, rolling hotfixes or virtual patching through network controls when a permanent fix isn’t immediately available. Organizations should also limit each account and system to the access it needs. Least privilege and network segmentation can make it harder for an attacker who steals credentials to move between systems or gain higher privileges.
Incident response plans need the same attention. Rehearsals should include scenarios involving compromised accounts and credentials, with teams practicing how to contain an attack when normal systems are unavailable. Recovery keys and emergency administrative accounts should be protected so the organization can regain control if its identity provider or domain is compromised.
The warning from Mythos is that powerful AI capabilities may spread beyond a limited preview, potentially with fewer safeguards. Security teams can’t base their readiness on access to one model or on assumptions about who will use similar tools. They need to make discovery and remediation faster, contain the effect of successful attacks and know how to recover.
The headline moment around Mythos may fade, but the operational challenge remains. Organizations that connect continuous exposure discovery with risk-based patching, strict access controls and practiced response will be better positioned to handle attacks that move at machine speed. Their measure of readiness is not how many flaws they can find, but how quickly they can reduce the danger those flaws create.