AI-Powered Cyberattacks in 2026: How AI Is Changing Vulnerability Discovery and Exploitation
Artificial intelligence is changing cybersecurity from both sides of the battlefield.
Security teams are using AI to analyse code, identify suspicious activity, prioritise vulnerabilities and accelerate investigations. At the same time, threat actors are experimenting with the same technology to improve reconnaissance, vulnerability research, exploit development and other parts of the attack lifecycle.
This is no longer purely theoretical.
In May 2026, Google Threat Intelligence Group (GTIG) reported that it had identified, for the first time, a threat actor using a zero-day exploit that Google believes was developed with AI assistance. According to GTIG, the criminal actor intended to use the vulnerability in a mass-exploitation operation.
That does not mean AI can automatically compromise any system.
It does mean the time and expertise required for some stages of vulnerability research may be changing—and defenders need to prepare for a faster vulnerability lifecycle.
The Traditional Vulnerability Race
Software vulnerabilities are nothing new.
Historically, vulnerability discovery often required security researchers to manually inspect source code, analyse binaries, fuzz applications, investigate unusual behaviour and reproduce crashes.
Once a vulnerability was discovered, several additional steps could follow:
Discovery → Analysis → Disclosure → Patch → Exploitation → Remediation
Attackers could perform similar research, but sophisticated vulnerability discovery and reliable exploit development traditionally required substantial technical expertise and time.
AI doesn't eliminate those requirements.
Instead, it can increasingly act as an accelerator.
AI systems can assist researchers with code analysis, pattern recognition, hypothesis generation, debugging and repetitive investigative tasks. This potentially allows humans—or increasingly autonomous agents—to examine more code in less time.
What AI Changes
The most important change may not be a completely new type of cyberattack.
It is speed and scale.
Google's analysis of 2025 zero-days predicted that AI would accelerate the competition between attackers and defenders during 2026, particularly through faster reconnaissance, vulnerability discovery and exploit development.
The difference can be understood like this:
Traditional vulnerability environmentAI-accelerated environmentManual reconnaissanceAutomated and agentic reconnaissanceHuman-heavy code reviewAI-assisted code analysisTime-intensive vulnerability researchAutomated vulnerability hypothesis generationManual exploit prototypingAI-assisted exploit developmentPeriodic security reviewIncreasingly continuous analysisCVSS-focused prioritisationExploitability + exposure + threat intelligenceReactive patch managementContinuous attack-surface reduction
The important point is that AI does not need to invent a completely new attack technique to create risk.
Making existing techniques faster, cheaper and easier to scale can itself change the threat landscape.
Evidence Emerging in 2026
Several developments this year illustrate where the technology is heading.
GTIG reported in May that it had identified a threat actor using a zero-day exploit it assesses was developed with AI. Its wider research also describes adversaries integrating generative AI into operational workflows rather than using it only as a basic coding assistant.
Mandiant has meanwhile been experimenting with agentic vulnerability discovery on the defensive side. Its recent work describes AI systems capable of analysing source code, developing vulnerability hypotheses and helping security researchers investigate complex security weaknesses.
These developments reveal something important:
The same capability that helps defenders discover vulnerabilities faster can potentially help attackers search for them faster.
This is a classic dual-use security problem.
From AI Assistant to AI Security Agent
Early generative AI cybersecurity use often looked like this:
Human asks question → AI produces answer → human performs next action.
Agentic systems can operate differently.
An AI agent may be able to perform a sequence of tasks, evaluate the results and determine what to do next.
For defensive vulnerability research, that could mean:
Inspect code → identify suspicious function → develop hypothesis → test hypothesis → analyse result → investigate another path.
Mandiant's research into agentic source-code review demonstrates why this model is attracting attention for vulnerability discovery.
This doesn't eliminate the need for skilled security researchers. Complex vulnerabilities, exploit reliability, environmental differences and false positives still require careful validation.
But AI can potentially remove significant amounts of repetitive work.
The Zero-Day Problem Becomes More Important
A zero-day vulnerability is a security flaw for which defenders may have little or no time to prepare before exploitation begins.
AI potentially makes this problem harder because vulnerability discovery itself can become more scalable.
Google's 2026 research warns of a period where AI can help identify weaknesses in existing software before AI-assisted defensive improvements have hardened the broader software ecosystem.
There is another uncomfortable reality.
Sometimes exploitation begins before a patch is available.
Google's July 2026 vulnerability-management research, referencing M-Trends 2026 data, reports a mean time-to-exploit of -7 days in its dataset—meaning exploitation was occurring, on average, before a patch became available for the vulnerabilities measured.
For defenders, that reinforces why vulnerability management cannot simply mean:
Wait for patch → install patch → problem solved.
Organisations also need detection, segmentation, exposure management and incident-response capabilities.
AI Does Not Make Every Vulnerability Exploitable
There is an important distinction between:
finding suspicious code
and
building a reliable real-world exploit.
An AI system may identify a potential weakness that turns out to be a false positive.
Even when a genuine vulnerability exists, exploitation can depend on architecture, configuration, authentication, operating-system protections and many other conditions.
Reliable exploitation may still require substantial human expertise.
Therefore, headlines suggesting that AI can simply “hack anything” should be treated carefully.
The evidence points toward capability acceleration, not unlimited autonomous hacking.
Vulnerability Prioritisation Must Also Change
Many organisations have thousands of vulnerabilities.
Trying to patch everything immediately is unrealistic.
Historically, teams often relied heavily on severity ratings such as CVSS.
Severity remains useful, but a critical question is:
Is this vulnerability actually being exploited?
That's where the U.S. Cybersecurity and Infrastructure Security Agency's Known Exploited Vulnerabilities (KEV) Catalog becomes particularly valuable.
CISA adds vulnerabilities to KEV when there is evidence of exploitation in the wild. The catalog continues to receive new additions; for example, CISA added another actively exploited vulnerability on August 21, 2026.
A stronger prioritisation model therefore considers:
Severity + active exploitation + internet exposure + asset importance + available mitigations
rather than severity alone.
What Defenders Should Do Now
AI doesn't make cybersecurity fundamentals obsolete.
It makes executing those fundamentals quickly and consistently more important.
1. Know What You Actually Own
You cannot protect infrastructure you don't know exists.
Maintain an accurate inventory covering:
servers
endpoints
cloud workloads
SaaS platforms
internet-facing applications
network appliances
software dependencies
privileged accounts
Unknown assets can become forgotten attack surfaces.
2. Prioritise Known Exploitation
Security teams should actively monitor the CISA KEV Catalog and relevant vendor advisories.
An internet-facing vulnerability with confirmed exploitation may deserve faster action than a higher-CVSS vulnerability on an isolated internal system.
CISA continues to recommend prioritising known exploited vulnerabilities, particularly on internet-facing infrastructure.
3. Reduce Internet Exposure
Every publicly reachable service increases attack surface.
Ask:
Does this system actually need to be publicly accessible?
Administrative interfaces, development systems, and unnecessary services should not be exposed simply because doing so is convenient.
4. Build an Emergency Patching Process
Normal monthly patch cycles may be too slow for actively exploited vulnerabilities.
Organisations should have an emergency process that allows critical patches or mitigations to be assessed, tested, and deployed rapidly.
5. Strengthen Identity Security
A vulnerability is not always the attacker's first entry point.
Credential theft, phishing, and session compromise remain major threats.
Use:
phishing-resistant MFA where practical
least privilege
separate administrator accounts
strong access controls
monitoring for suspicious authentication
6. Improve Detection
Assume prevention will occasionally fail.
Collect and protect meaningful logs from:
identity providers
endpoints
servers
firewalls
cloud platforms
critical applications
Detection gives defenders another opportunity to stop an intrusion after initial access.
7. Secure Software Earlier
Vulnerability management shouldn't begin after deployment.
NIST's Secure Software Development Framework encourages organisations to integrate secure-development practices into the software development lifecycle rather than treating security as a final-stage activity.
AI-assisted defensive code review may strengthen this approach further.
AI Is Also Strengthening Defenders
It would be misleading to discuss AI only as an attacker advantage.
Defenders have access to the same broad technological shift.
AI can help security teams:
analyse large codebases
investigate alerts
correlate threat intelligence
identify suspicious behaviour
summarise incidents
prioritise vulnerabilities
assist malware analysis
accelerate security testing
NIST's emerging Cyber AI work explicitly considers both AI-enabled cyber attacks and AI-enabled cyber defence, reflecting this dual-sided transformation.
The future therefore isn't simply:
AI attackers vs human defenders.
It is increasingly:
AI-assisted attackers vs AI-assisted defenders.
The organisations that integrate automation responsibly may be better positioned to keep pace.
What Happens Next?
The next stage of AI-enabled cybersecurity is likely to involve increasingly capable agents.
Instead of assisting with a single task, these systems may perform longer sequences of vulnerability research, defensive analysis and security operations.
That creates opportunities and risks.
AI-assisted development could eventually result in software that is substantially harder to exploit because vulnerabilities are detected during development.
But the transition period matters.
Legacy systems, unpatched applications and exposed infrastructure still exist today, while vulnerability-discovery technology is advancing rapidly.
That mismatch may create a particularly challenging period for defenders.
Final Thoughts
AI is not replacing cybersecurity fundamentals.
It is compressing the timeline around them.
Vulnerability discovery can become faster. Reconnaissance can become more scalable. Security analysis can become more automated. Defenders may have less time between disclosure, exploitation, and required action.
But AI also gives security teams powerful new capabilities.
The organisations best positioned for this environment will not be those chasing every AI headline. They will be the ones combining modern automation with strong fundamentals:
Know your assets. Reduce exposure. Prioritise real exploitation. Patch quickly. Protect identities. Monitor continuously. Prepare for incidents.
The vulnerability race is accelerating.
Defenders need to accelerate with it.
References
For the article's References field, use these primary sources:
