AI-generated malware is becoming easier for criminals to produce, scale and adapt. Yet its most recognizable clues may be surprisingly ordinary: tidy comments, uniform logging messages and emojis.
In an interview with Help Net Security, Lindsay Kaye, vice president of threat intelligence at HUMAN Security, described what AI-assisted code can look like, why the window for patching vulnerabilities is contracting, and why knowing an AI model was involved rarely helps defenders decide what to do next.
- Malware written entirely with AI has not appeared in Kaye’s own work, to her knowledge.
- Tidy comments, consistent logging text and emojis can provide clues, but none proves AI authorship.
- Decompilation and obfuscation can erase much of that context.
- AI can help criminals find vulnerabilities and develop exploits more quickly.
- Kaye’s team requires human review before releasing AI-assisted conclusions.
What AI-generated malware may look like
Language-model-generated code often appears planned and structured, following familiar design patterns. The stylistic clues Kaye encounters include unusually orderly comments, uniform logging messages and emojis. These features can make a script look polished, particularly when compared with inconsistent or hastily written code.
That appearance is evidence, not attribution. Not every script with these characteristics is AI-generated, and legitimate human-written code can follow the same conventions. A defensible assessment therefore considers several clues together rather than treating one tidy comment as a verdict.
The clues also become less useful once investigators move to decompiled or disassembled code. Decompilers cannot perfectly rebuild C/C++-like source, so context disappears. Obfuscation can remove or conceal the comments, logging strings and formatting that initially attracted attention. Investigators may recover what the program does without determining how it was produced.
That limitation leads to the central defensive lesson: provenance is context, while behavior is actionable. An “AI-generated” label does not tell a security team whether to block a file, investigate an account or prioritize a patch.
Where AI changes the criminal economics
AI’s clearest effect is not necessarily a dramatic fall in the price of stolen accounts. Kaye said she has not seen drastic pricing changes she can directly attribute to AI; pricing remains largely connected to supply and demand.
The more significant gains are scalability and efficiency. AI makes it easier to assemble scripts for a particular purpose and helps threat actors improve the acquisition and verification of account-related assets. Even without cheaper criminal goods, the ability to conduct more of the work at scale can increase pressure on defenders.
Adaptability may be even more consequential. Kaye highlighted PROMPTSTEAL as an example of malware using AI to alter its behavior during execution. This moves beyond faster script creation: the malware can potentially change how it operates while an attack is underway.
Criminals benefit from that combination of speed, scale and flexibility. Defenders, meanwhile, must make sense of more output while operating with less patching time.
Why the patching window is shrinking
The increased availability of AI has made finding security flaws more accessible and easier to accomplish at scale. It has also created a quicker and potentially easier path from vulnerability to exploitation.
Kaye cited the PHP2Shell remote-code-execution vulnerabilities as an example. If threat actors had widely exploited them rather than researchers disclosing them, the potential damage could have been exceptional, particularly because a patch was not yet available.
Patching has never been simple. Security teams must manage dependency versions, ensure a patch will not break other parts of their environment, and account for downtime when deploying it. Those requirements can stretch the time between discovery and remediation.
Public proof-of-concept exploits add further pressure. There is no universal patch window that suits every organization, but AI-enabled threat actors are accelerating the timeline defenders must respond to. The asymmetry is straightforward: attackers can move faster while defenders remain responsible for testing safety and operational continuity.
Why human judgment remains in the loop
Kaye’s team uses AI to augment threat-intelligence workflows. The technology can help summarize information, answer questions about collected data and extract indicators of compromise from reports. It can also support the creation of detection rules used to hunt for malware related to samples of interest.
Those outputs are starting points rather than final findings. Kaye said the results cannot be shared as they are. A person must confirm that the conclusions are correct and review them for additional insight the model may have missed.
The same principle applies to malware attribution. A hash or domain is specific and reasonably durable, whereas knowing that a model was involved is neither. More granular indicators such as prompt patterns can become obsolete as models change. Kaye’s preferred approach is to document techniques in the frameworks detection engineers already use, with downstream teams contributing what they observe about AI-enabled activity.
A decision checklist for defenders
Use the following checklist when AI-generated code or AI-assisted intelligence enters an investigation:
- Separate style from behavior:** Record tidy comments, uniform logging messages or emojis as context, not proof.
- Ask what the activity does:** Base investigation and response decisions on the behavior observed on the target.
- Account for lost context:** Expect decompilation and obfuscation to weaken stylistic indicators.
- Prioritize exposure:** When vulnerabilities become public, determine whether affected software is in use and whether proof-of-concept exploitation changes the urgency.
- Map patch constraints:** Identify dependency, compatibility and downtime issues that could delay remediation.
- Validate AI-assisted findings:** Require a human analyst to confirm conclusions against the source data and add relevant context.
- Document techniques where defenders use them:** Add observed methods to existing detection frameworks rather than relying on a separate “AI abuse” alert.
What to watch is whether AI-assisted techniques become better documented and more consistently integrated into detection engineering. Kaye is scheduled to participate in the Nullcon Berlin panel discussion “Agents Gone Wild” on November 5, 15:30–16:15.
Key takeaways
- Treat stylistic clues as supporting evidence, never as definitive AI attribution.
- Prioritize observed behavior when deciding whether to investigate, block or patch.
- Recognize that AI can shorten vulnerability discovery, exploit development and patching timelines.
- Document malicious techniques in existing detection frameworks.
- Keep a human analyst responsible for conclusions that trigger consequential action.
Frequently asked questions
Can you tell whether malware was written by AI?
Not with certainty from comments, logging messages or emojis. These are clues rather than proof, and decompilation and obfuscation can erase them. Human-written code can also exhibit the same characteristics.
How does AI make patching more difficult?
It helps threat actors find vulnerabilities and reach exploitability more quickly, while defenders still need to test dependencies, check compatibility and account for downtime. The result is a narrower window for safe remediation.
Should defenders create a separate alert for “AI abuse”?
Kaye questioned whether that is the most useful unit. Knowing a model was involved does not reliably indicate what defenders should do, while the activity on the target provides a more actionable basis for response.
How is Kaye’s team using AI defensively?
The team uses AI to summarize and query collected data, extract indicators of compromise and assist with detection-rule creation. Every conclusion receives human confirmation and review before release.