What Happens When AI Creates an Attack You’ve Never Seen?

What happens when the attacker discovers a vulnerability your software vendor does not even know exists?

That is becoming one of the most important cybersecurity questions business leaders need to consider.

For years, the discussion around artificial intelligence and cybercrime focused largely on better phishing emails, faster malware development, and helping less-skilled attackers become more capable.

That conversation has changed.

OpenAI now warns that threat actors will increasingly use AI to conduct cyberattacks at “unprecedented speed and scale,” including in fully autonomous ways. Axios recently reported that OpenAI, Anthropic, Microsoft, AWS, and more than 100 organizations are warning that defenders have a rapidly narrowing window to prepare for increasingly capable AI-driven attacks. (OpenAI)

But speed is only part of the problem.

The more serious issue may be that AI can help create attacks that are unknown, adaptive, and much harder for traditional detection tools to recognize before damage occurs.

So what exactly is changing?

AI gives attackers something they have never had at this scale: the ability to continuously analyze software, systems, credentials, and possible attack paths at machine speed.

The defensive side already shows what this technology can do.

Anthropic reported that its AI-assisted vulnerability research had disclosed 2,300 vulnerabilities across 392 open-source projects as of August 26, 2026. Only 421 were known to have been patched at that point. (Anthropic Red)

Microsoft has demonstrated similar capabilities. Its agentic security system, known as MDASH, helped researchers discover 16 previously unknown vulnerabilities in Windows networking and authentication components, including four critical remote-code-execution flaws. Microsoft described the results here. (Microsoft)

These are defensive uses of AI.

But the implication is difficult to ignore.

If AI can help defenders discover vulnerabilities software vendors did not previously know existed, attackers can potentially use similar capabilities to search for unknown weaknesses.

There is no patch for a vulnerability nobody knows exists.

And there may be no detection rule for an attack technique nobody has seen before.

Why does this create such a problem for detection?

Most modern endpoint-security strategies still depend heavily on identifying something suspicious.

That might be known malware, abnormal behavior, a malicious file, an indicator of compromise, suspicious process activity, or enough unusual telemetry to trigger an alert.

AI-assisted attackers do not necessarily need to operate that way.

They can potentially abuse valid credentials, PowerShell, scripting engines, browser sessions, remote-management tools, trusted applications, and other technologies already present inside the environment.

This is often called living off the land.

Instead of introducing an obviously malicious program, an attacker can ask a much simpler question:

What is this computer already allowed to do?

AI can then help evaluate those options, choose an attack path, observe the results, and adapt.

That is why the next generation of attacks may be difficult to recognize. The attacker does not necessarily need to remain invisible forever.

It only needs to remain invisible long enough to succeed.

This is what I call the Detection Gap

The Detection Gap is the time between when malicious activity begins and when security systems understand enough about that activity to identify it as an attack.

AI threatens to make that gap far more dangerous.

Attackers may increasingly be able to move through stages such as:

discover → analyze → exploit → observe → adapt → continue

without waiting for a human operator to perform every step.

Axios reported in August that organizations may have only a short window before models capable of end-to-end autonomous cyberattacks become available to malicious actors. (Axios)

That changes the timing problem dramatically.

Detection can be accurate and still be too late.

But aren’t EDR and Detect and Respond designed for this?

Detection and response still matter.

The problem is relying on them as the final barrier between an attacker and business damage.

The Verizon 2025 Data Breach Investigations Report found that exploitation of vulnerabilities as an initial access method increased 34% year over year. Ransomware was present in 44% of the breaches Verizon reviewed, up from 32% the previous year. (Verizon)

Verizon also found that credential abuse remained the most common known initial-access vector, appearing in 22% of the breaches where the initial vector was known. (Verizon)

Those are exactly the kinds of attack paths AI can accelerate.

The concern is not that EDR suddenly becomes useless.

The concern is that Detect and Respond assumes there will be enough time to detect, understand, and respond before the attacker accomplishes the objective.

AI is compressing that timeline.

What does this mean for the business?

The financial consequences of cyber incidents are already substantial.

According to IBM's 2025 Cost of a Data Breach Report, the global average cost of a data breach was $4.44 million, while the average U.S. breach reached a record $10.22 million. (IBM)

Those costs can include:

operational downtime
lost employee productivity
interrupted customer service
forensic investigation
recovery and restoration expenses
regulatory and legal exposure
lost revenue
reputational damage

AI does not create those consequences.

It potentially allows attackers to reach them faster.

So what needs to change?

Cybersecurity strategies need to assume that some attacks will not be recognized in time.

That means changing part of the security conversation from:

“Can we identify whether this activity is malicious?”

to:

“Should this application or process be allowed to perform this action at all?”

That is where Isolation and Containment become increasingly important.

The objective is to prevent unauthorized execution, restrict what applications can access, limit attacker movement, reduce the blast radius, and prevent encryption or destructive activity from occurring in the first place.

AppGuard is a proven endpoint protection solution with a more than 12-year track record focused on prevention through Isolation and Containment.

The goal is not to eliminate detection.

It is to stop making detection the only thing standing between an unknown attack and business damage.

What Should Businesses Do Next?

Business leaders should begin operating under the assumption that detection will occasionally fail or arrive late.

Add prevention layers that restrict unauthorized application behavior. Reduce unnecessary endpoint execution freedom. Segment critical systems. Review third-party and remote-access pathways. Protect credentials and browser sessions. Test what happens when EDR does not generate an alert. And maintain incident-response and recovery plans that have actually been exercised.

Most importantly, ask your IT provider or cybersecurity team a different question.

Do not ask only:

“Will our security detect the next attack?”

Ask:

“If it does not, what stops the attack from succeeding?”

That question becomes much more important as cyberattacks move from human-speed operations toward autonomous, adaptive attacks operating at machine speed.

Business owners who want to better understand how prevention-first security can stop attacks before damage occurs should talk with CHIPS about how AppGuard can help prevent incidents like this through Isolation and Containment.

Additional Source Reading

Tags:

Post by Tony Chiappetta
August 31, 2026