From "AI That Finds Bugs" to "AI That Constructs Attack Paths"… Defense Industry Timeline Accelerated
The era when AI writes code well has already arrived. The more important question now is this: how well can AI find the weaknesses in code, and how quickly can it convert those weaknesses into actual attack possibilities?
The cybersecurity evaluation results of Claude Mythos Preview published by Anthropic send a very strong signal regarding this question. In announcing Claude Mythos Preview on April 7, 2026, Anthropic stated that this model demonstrated powerful cybersecurity capabilities capable of identifying zero-day vulnerabilities in major operating systems and web browsers, with some even verifying actual attack possibilities.
Alongside this announcement, Anthropic launched Project Glasswing. This is a defense-focused project aimed at providing Mythos Preview to limited partners and key open-source developers to find and fix vulnerabilities in important software before attackers do. Anthropic stated it would not publicly release Mythos Preview, acknowledging that this model's capabilities could have significant impact on both defense and offense.
Claude Mythos Preview and Project Glasswing
Anthropic introduced Claude Mythos Preview as a general-purpose frontier model. The key point is that this model was not specially trained solely for cybersecurity. As code understanding, reasoning, and autonomous task capabilities improved overall, vulnerability detection and exploit construction capabilities also dramatically increased as a result.
According to Anthropic's technical report, Mythos Preview demonstrated clearly different performance levels from existing Claude Opus 4.6 in internal evaluations. Opus 4.6 was strong in vulnerability identification and correction, but its success rate in autonomous exploit development was close to 0%. In contrast, Mythos Preview reportedly constructed working exploits at a much higher rate in experiments related to the Firefox JavaScript engine.
Anthropic created Project Glasswing to provide this capability to defenders first rather than spreading it to attackers. The purpose of this project is to support key industry partners and open-source developers in pre-checking and fixing important software.
AI Has Moved from 'Vulnerability Discovery' to 'Attack Possibility Verification'
Until now, expectations for AI-based security tools were primarily in vulnerability detection. The method was to read code, find dangerous patterns, flag vulnerable function usage, and suggest patch drafts.
However, the reason Mythos Preview's announcement is shocking is that it went one step further. Anthropic stated that this model went beyond simply saying "there may be a bug here" to constructing exploits that verify actual attack possibilities.
Of course, this point is extremely sensitive. Actual attack procedures or reproducible technical details should not be publicly disseminated. However, the industrial significance is clear.
If AI finds vulnerabilities faster, the time until patching shortens. Within that short time, attackers can also move. Therefore defenders must find them faster and fix them faster.
The core competitive factor in cybersecurity is shifting from 'who discovers first' to 'who fixes and deploys first.'
Cases Presented by Anthropic: Old Bugs Are Reopened Before AI
The Anthropic report explains that Mythos Preview discovered vulnerabilities in various domains including OpenBSD, FFmpeg, FreeBSD, Linux kernel, web browsers, and cryptographic libraries. In particular, the report mentions cases such as a 27-year-old OpenBSD bug, a 16-year-old FFmpeg vulnerability, and a remote code execution vulnerability in the FreeBSD NFS server.
The core of these cases is that the belief that "old code is safe" is being shaken. Even in code that has been in use for decades and reviewed by countless developers, fuzzers, and security researchers, subtle logic errors and memory safety issues may still remain.
Anthropic explains in particular that Mythos Preview operated by repeatedly reading large codebases, hypothesizing suspicious points, verifying in execution environments, and revising again. This resembles the method of human researchers but differs in being able to explore many more files and paths without fatigue.
What is important here is not that AI invented entirely new hacking principles. Rather, the report explains that it combines known techniques from existing security research such as ROP, JIT heap spray, and KASLR bypass. The problem is that this combination can be automated and repeated at scale.
External Evaluations Also Emerged: "Step Change in Cyber Capabilities"
The UK AI Security Institute also evaluated Claude Mythos Preview's cyber capabilities. AISI stated that Mythos Preview showed clearly improved performance over previous frontier models in CTF challenges and multi-stage cyberattack simulations.
The security industry also interprets this announcement as a cybersecurity inflection point. Bishop Fox analyzed that Mythos Preview narrows the gap between vulnerability discovery and exploit possibility assessment, and will pressure existing security organizations' vulnerability management, patching, and verification systems.
The Mozilla case also draws attention. According to Business Insider reports, Mozilla used Claude Mythos Preview to discover and patch hundreds of Firefox security issues, with many reportedly directly identified by Mythos Preview. This demonstrates that AI-based vulnerability detection is moving beyond the laboratory and into actual large-scale software development processes.
Tool of Defense or Automation of Attack?
The greatest dilemma of Mythos Preview is its dual-use nature. The same capability can be a powerful security inspection tool for defenders, but can become an automated vulnerability discovery and attack tool for attackers.
Anthropic acknowledges this point. The report projects that in the long term, powerful language models will bring greater benefits to defenders. However, it warns that in the short-term transition period, attackers may gain advantages first. That's why Mythos Preview was not publicly released but provided first to limited defense partners.
This judgment sets new standards for AI model release strategies. The more powerful a model is, the harder it becomes to distribute widely simply because 'performance is good.' Especially in areas like cybersecurity where attack and defense share the same technical foundation, access control, use purpose verification, output restrictions, and audit systems are also necessary.
This means AI safety issues are not limited to misinformation or bias. AI safety is now expanding to encompass vulnerability discovery capabilities, attack automation possibilities, and control over model access.
Patch Cycles Can No Longer Be Leisurely
Anthropic offers several practical recommendations to defenders in the report. The core is clear: reduce patch cycles.
Previously, there was a certain amount of time between a disclosed vulnerability and actual attack code. This was because specialized researchers needed to analyze patches, reconstruct vulnerabilities, and create attack code.
However, Mythos Preview-level models can verify attack possibilities much faster using only published CVEs and patch commits. Anthropic explains that N-day vulnerabilities — already known but not yet patched on many systems — can become more dangerous.
This requires direct changes in corporate security operations. Security patches should no longer be classified as monthly routine work but as emergency response. Automatic updates should be expanded in possible domains. Open-source dependency updates should be viewed not as simple maintenance but as attack surface reduction. Systems for AI to first classify the authenticity and severity of vulnerability reports are also needed.
In other words, security organizations in the AI era must shift from "reviewing when vulnerabilities are discovered" to "operating with the premise that large volumes of vulnerabilities can appear at any time."
Corporate and Government Risks: The Era of Vulnerability Floods Arrives
The future Mythos Preview has shown is not simply a future where attacks become stronger. More precisely, it is a future where the volume of vulnerability discoveries explodes.
Until now, vulnerabilities were found by scarce specialized personnel. However, if AI can read large codebases in parallel, repeatedly verify suspicious points, and write reports, the situation changes. Security teams will receive more reports, need to evaluate more patches, and handle more exceptional situations.
The problem is that not all organizations are prepared to keep up with that pace. Legacy systems, old firmware, discontinued products, code inherited through mergers and acquisitions, and internal applications whose developers have departed are particularly vulnerable.
Anthropic suggests that defenders should review vulnerability disclosure policies and prepare emergency mitigation strategies for legacy software and hardware in advance.
This is also an important message for governments and critical infrastructure operators. Finance, telecommunications, energy, transportation, defense, and healthcare systems have large-scale software dependencies. If AI increases vulnerability discovery speed, attackers can also utilize that speed. If defense systems are slow, disclosed vulnerabilities immediately become attack opportunities.
Cybersecurity Moves to 'AI vs. AI' Competition
The essence of this announcement is not "Anthropic has built a powerful security AI." The larger change is that the operating method of cybersecurity is changing.
Going forward, attackers will use AI to read code, analyze patches, and combine vulnerabilities. Defenders will also use AI to inspect code, summarize logs, detect breach indicators, and prioritize patches.
Cybersecurity is increasingly becoming an AI vs. AI speed competition.
What is important is not a single model's performance but the organization's response system. Having AI security tools is meaningless if patch approval procedures are slow. Finding vulnerabilities does not reduce risk if the development team has no capacity to fix them. Analyzing logs does not help attackers if decision makers cannot determine response priorities.
Therefore, cybersecurity competitiveness in the AI era is redefined by three factors: how quickly vulnerabilities can be found, how accurately the actual risk level of those vulnerabilities can be judged, and how quickly patches and mitigation measures can be deployed.
Mythos Preview has accelerated the timeline of these three questions.
The Future of AI Security Can Be Safer, But the Transition Is Dangerous
Claude Mythos Preview brings both optimism and anxiety to cybersecurity simultaneously.
An optimistic future also exists. If AI finds numerous vulnerabilities in advance, proposes patch code, and reduces risk in legacy systems, the software ecosystem may become safer in the long term. Anthropic also projects that defenders will gain greater benefits in the long run.
However, the transition period is dangerous. If AI can quickly verify attack possibilities, organizations with delayed patches may be attacked in shorter timeframes. The buffer time between vulnerability disclosure and actual exploitation shrinks. Security team workloads increase, and conventional manual-centered vulnerability management systems will face pressure.
The message this announcement delivers is clear.
AI has now become not just a tool for writing code but a tool for finding code's weaknesses. And it has begun to judge whether those weaknesses can become actual attacks.
Therefore, the question for cybersecurity also changes.
Not "Is AI dangerous?" but "Are we ready to fix vulnerabilities at the pace AI finds them?"
Claude Mythos Preview is a model name, but its significance is much larger. The equilibrium point of cybersecurity has begun to shift. Defenders must now move faster than that shift.
