The trigger was small enough to miss inside a much larger cybercrime investigation.
Google says someone connected to the TeamPCP cybercrime group used an AI tool to help develop a previously unknown exploit targeting widely used login software. The exploit could bypass two-factor authentication.
Google obtained the code, tested it and found that, after several adjustments, it worked. The software developer was notified and patched the vulnerability. Google described the case earlier this year as the first time its threat-intelligence team had identified a threat actor using a zero-day exploit it believed had been developed with AI assistance. WIRED
That does not mean AI can now autonomously discover and weaponize every zero-day.
It means something more important has started.
AI has entered the vulnerability race on both sides.
Attackers are using it to search, code and test.
Defenders are using it to scan enormous codebases, identify weaknesses and generate fixes.
The advantage may increasingly belong to whichever side can move from discovery to action faster.
Cybersecurity has traditionally operated around a human-speed cycle.
A researcher finds a vulnerability.
Someone validates it.
A vendor investigates.
A patch is written.
Customers are notified.
Attackers reverse-engineer the fix or independently discover the weakness.
Security teams deploy the update.
That cycle already moves too slowly.
Google’s Mandiant unit says the mean time-to-exploit has fallen to negative seven days — meaning vulnerabilities are, on average in its dataset, being exploited before a patch becomes available. Google Cloud
AI threatens to compress that cycle further.
Google Threat Intelligence Group reported in September that monthly vulnerability disclosures rose from 5,045 in January 2026 to 10,740 in August. Vulnerabilities observed being exploited increased from an average of 10.5 per month in 2025 to 18 per month in the first eight months of 2026. Zero-day exploitation rose more modestly, from about eight per month to eleven. Google Cloud
Those figures do not prove AI caused every increase. Google itself cautions that disclosure practices, vendor activity and other factors affect the numbers.
But its researchers say AI is already changing the vulnerability landscape — including the kinds of flaws being discovered.
That matters because AI does not simply make an existing researcher type faster.
It can potentially search thousands of repositories simultaneously, compare patterns across enormous codebases, generate hypotheses, test them, reject failures and continue.
The same automation applies to defense.
The emerging contest is therefore not:
AI attackers versus human defenders.
It is:
AI-assisted attackers versus AI-assisted defenders.
Finding valuable software flaws has historically required scarce expertise and significant time.
AI changes the economics.
Anthropic says organizations participating in Project Glasswing found at least 129,000 verified software vulnerabilities between April and July 2026, with more than 33,000 eventually rated high or critical. Anthropic’s own automated open-source scanning identified another 5,500 verified vulnerabilities through October. These are Anthropic-reported figures, but their scale illustrates what happens when advanced models can repeatedly inspect software. Anthropic
More vulnerabilities become discoverable because the cost of looking falls.
That benefits defenders.
It also expands the hunting ground available to attackers.
Discovering a weakness and turning it into something usable are different jobs.
The TeamPCP incident matters because AI appears to have crossed that boundary.
Google says the group was not merely asking a chatbot to explain security concepts. An AI tool had been used in developing exploit code against a previously unknown authentication flaw. WIRED
The immediate consequence is not autonomous hacking.
It is faster iteration.
An attacker can ask AI to analyze code, propose attack paths, modify an exploit after a failed test and automate repetitive development work that once consumed specialist time.
That can make moderately skilled operators more capable and highly skilled operators faster.
The defensive response is already moving in the same direction.
Google says it is embedding AI vulnerability-scanning agents into its own software-development lifecycle, continuously reviewing changes across hundreds of millions of lines of infrastructure code. The company says these systems are preventing hundreds of vulnerabilities each month from reaching production. Google Cloud
This is the other half of the signal.
AI can shorten the attacker’s clock.
It can also shorten the defender’s clock.
Security increasingly becomes a contest over who closes the loop first.
Machine discovery can produce findings much faster than humans can investigate them.
Anthropic’s coordinated vulnerability-disclosure program illustrates the problem. Its systems generated tens of thousands of candidate findings, but external security firms and human reviewers still have to validate them, determine severity and contact maintainers. Anthropic
The bottleneck begins moving.
Yesterday it was finding enough vulnerabilities.
Tomorrow it may be deciding which of thousands of machine-discovered vulnerabilities actually matter first.
The traditional model of scheduled software updates becomes increasingly uncomfortable if vulnerability discovery operates continuously.
Enterprises may have to move toward:
continuous detection → automated prioritization → automated testing → automated remediation.
Humans remain responsible for high-risk decisions, but many routine fixes may move directly through machine-assisted pipelines.
Security becomes less like periodic maintenance and more like a constantly running immune system.
A zero-day is valuable partly because the defender does not know it exists.
If AI systems become better at independently rediscovering vulnerabilities, the useful life of a secret exploit could shrink.
That changes attacker economics.
A vulnerability that might once have remained hidden for months may be independently discovered by a defensive model much sooner.
Ironically, faster offensive AI could eventually force faster defensive AI — reducing the period during which some vulnerabilities remain secret.
Advanced cyber operations have traditionally required significant expertise.
AI will not erase that requirement, particularly for sophisticated attacks.
But it can lower portions of the skill barrier.
Someone who understands the objective but lacks deep experience in one programming language, operating system or exploit technique can increasingly use a model to fill knowledge gaps.
That expands the number of actors capable of attempting technically serious attacks.
The highest-end operators benefit too because AI removes routine work from people who already possess the expertise.
If machines can inspect software at scale, poorly maintained code becomes more expensive.
Old dependencies.
Unpatched libraries.
Exposed APIs.
Weak authentication.
Hard-coded credentials.
Abandoned open-source packages.
Those weaknesses become easier for both sides to locate.
Companies that treat software maintenance as back-office housekeeping may discover that it has become part of frontline security.
There is another problem.
What happens when AI discovers vulnerabilities faster than maintainers can patch them?
Security researchers cannot responsibly dump thousands of serious vulnerabilities into public databases before vendors have time to respond.
But withholding too much information also creates risk.
The vulnerability-disclosure system itself may need automation for prioritization, vendor notification, patch validation and coordinated release.
The machines may eventually discover vulnerabilities faster than today’s human institutions can process them.
Security platforms with strong AI automation gain value because companies will need continuous vulnerability discovery, prioritization and remediation rather than occasional scanning.
Large software vendors with deep engineering resources can deploy defensive agents across entire codebases before vulnerabilities reach production.
Verified security researchers and red teams gain increasingly capable tools. Anthropic’s newly expanded Cyber Verification Program explicitly gives vetted defenders access to stronger cyber capabilities while maintaining greater restrictions for general users. Anthropic
Cloud providers and infrastructure companies become central because they can see activity across enormous environments and deploy defenses at machine scale.
Well-maintained open-source projects may benefit from AI finding flaws that volunteer maintainers would never have had enough time to discover manually — assuming they have enough capacity to patch what the machines find.
Organizations with slow patch cycles become increasingly exposed.
A company taking 30, 60 or 90 days to remediate serious vulnerabilities will struggle in an environment where discovery and exploit development accelerate.
Legacy software becomes more dangerous because machines can repeatedly probe code that organizations have avoided touching for years.
Security teams built around manual triage face capacity problems. Humans cannot individually investigate every machine-generated alert if discovery scales by orders of magnitude.
Attackers relying on long-lived private exploits could eventually lose some advantage if defensive AI becomes capable of independently rediscovering those same weaknesses faster.
And companies assuming their existing security process is fast enough may discover that the relevant comparison is no longer their response time against another company’s.
It is their response time against a machine.
The number to watch is not simply how many vulnerabilities AI finds.
Watch the time between each stage.
Discovery to validation.
Validation to exploit.
Discovery to vendor notification.
Notification to patch.
Patch to deployment.
Those intervals will tell us whether AI is genuinely changing cybersecurity.
Google is already warning that organizations need to move away from indiscriminate mass patching toward intelligence-driven prioritization and automated agentic remediation as vulnerability volumes rise. Google Cloud
Expect model companies to separate cyber access more aggressively as well.
Anthropic’s three-tier system is an early example: defensive researchers receive one level of capability, authorized red teams receive another, and highly vetted organizations protecting critical systems receive deeper access. Anthropic
Expect more companies to build dedicated vulnerability agents.
Expect security products to integrate automatic patch generation.
Expect attackers to incorporate models directly into exploit-development pipelines.
And expect regulators and software vendors to pay closer attention to what happens when an AI system discovers a serious vulnerability that no human knew existed.
The critical competition will be speed.
Not raw model intelligence by itself.
Which side detects first?
Which side validates first?
Which side acts first?
The TeamPCP zero-day is important because it provides a concrete glimpse of a transition that security researchers have been warning about.
AI is moving beyond phishing emails and malware assistance.
It is beginning to participate in the deeper technical cycle of finding, understanding and exploiting previously unknown software weaknesses.
At the same time, Google, Anthropic and other security organizations are deploying increasingly capable AI systems to search for those same vulnerabilities and fix them.
That means the long-term outcome is not predetermined.
AI could make software dramatically less secure if attackers consistently move first.
It could make software dramatically more secure if defensive systems continuously discover and repair weaknesses before they become useful weapons.
Most likely, both happen simultaneously.
The important shift is the clock.
Cybersecurity has always been a race between discovery and response.
AI is starting to remove the humans from the slowest parts of that race.
And when both sides operate at machine speed, a zero-day may not remain a zero-day for very long.
Signal Strength: 9.8/10
Confidence Level: High
Time Horizon: Immediate → 1–3 Years
Subscribe to follow the Trend newsletter and more.
Pass along insights or story ideas on AI, startups, and business. Focused on signal over noise, impact over headlines. Facts. Trends. Consequences. Always.
Buy Grey Ghost a Coffee
