It’s no secret that AI is being picked up by security defenders and malicious actors alike to transform their degrees of cyber readiness. On one hand, we have defenders using AI-powered detection engines to hasten incident response while, on the other, we have attackers multiplying their efforts, thanks to newly unlocked automation-sourced efficiencies.

Who wins this arms race may come down to early warnings and preemptive efforts, none of which are possible without substantial intelligence gathering on both sides. The good news, though, is that SOCs have one thing bad actors don’t: the promise of deterrence.

Key points of this article:

  • The attribution game is changing, and the old signals may not be as reliable as they once seemed.
  • AI is not only adding noise to the threat landscape; it may also create new patterns for defenders to study.
  • For SOCs, the real advantage may come less from certainty and more from speed, context, and correlation.
  • Deterrence in the AI era could depend on making attacks harder to sustain, even when the attackers’ identities remain unclear.
  • The next test for defenders may be turning AI from a source of uncertainty into a practical tool for faster, sharper responses.

Attribution: The key to victory

What makes good intelligence? ESET Director of Threat Research Jean-Ian Boutin’s answer would be all the behind-the-scenes work, the “heavy lifting of processing, enriching, correlating, clustering, and dissecting the samples that appear in telemetry.”

But how does this help SOC teams? And what does this have to do with attribution?

In essence, it is that same background work that builds intelligence that helps SOCs attribute attacks against their protected environments with higher precision. In a past story by ESET’s resident security story expert, James Rodewald, it was access to multiple forms of intelligence that led them to pinpoint an attacker, which, in their case, was Mallox ransomware.

Thus, understanding whether an organization faces a financially motivated cybercriminal (like a ransomware actor) or a state-aligned attacker can shape response priorities, strategic investments, and future threat hunting efforts.

The challenge with proper attribution

But it’s not as easy as it sounds. Attribution has never been simple. Both researchers and analysts might find it challenging as attackers often:

  • Reuse and share tools (or purchase them from MaaS operators)
  • Hide behind infrastructure (like legitimate cloud services or ORB networks)
  • Mimic other groups (two actors from Russia might have similar MOs), either inadvertently or conversely for introducing false flags 
  • Operate across borders, providing a degree of plausible deniability1, especially to state-aligned groups

On top of this are different research perspectives, where TTPs might overlap, but small differences in identification could equate to diverse attribution. At that point, it’s up to the threat hunters to connect the dots, especially when they have access to multiple different CTI feeds at once (which an SOC usually does).

AI enters the attribution picture

Here’s where things turn even murkier. AI is already affecting some of the signals analysts rely on for attribution.

Case in point: PromptLock. This discovery by ESET is especially interesting since it was first found on VirusTotal, which is not a surprise. But the fact that it very well could have been the vestige of an actual attack, as opposed to research, is where the lines get blurrier. 

But this is not limited to malware. In phishing, AI threatens to erase another familiar category of attribution clues: language (grammar, style, references). Gen AI reduces many of those signals by producing more convincing content—at scale.  The same is true for fake online personas, adding a bit of masking to the attackers’ soup.

Where things go gray

Why does this matter? If AI can already generate phishing lures, create deepfake-enabled scams, and assist with coding tasks, it is reasonable to expect that it will play an increasingly active role across the attack chain. As systems become more agentic, researchers may find themselves attributing campaigns in which a growing share of activity is automated rather than directly performed by a human operator. ESET has already identified thousands of seemingly benign skills2 that, upon closer inspection, exhibited malicious capabilities.

There was also PromptSpy, which used GenAI to achieve persistence. But what if a whole attack was orchestrated by AI? Recent incidents reported by OpenAI and Anthropic suggest that their models are capable of escaping sandbox restrictions to break into external systems. This further raises the stakes of asking “who” an attacker truly is. Is it the AI, or whoever set the attack in motion? And where does that leave the model provider? This effectively adds another layer of deniability between attackers and victims, which could significantly complicate law enforcement action.

But that’s not where the story ends. Deniability is one thing, but the increase in scale will undoubtedly generate more data points to analyze. In other words, AI may blur the fingerprints, but it also leaves behind more smudges to study.

Thus, the same volume that makes attribution harder for a human analyst may become useful when SOCs can correlate those data points with the help of …drumbeat … AI. And it isn’t such a far-fetched idea, either.

The defender’s AI advantage

First things first: AI isn’t meant to replace analysts. If it can enable bad actors to operate at scale, it can also ingest these data overflows and send them back into security dashboards, correlating signals, speeding up triage, detecting anomalies, and ultimately shortening investigation times.

More importantly, AI can help solve some of the attribution challenges discussed earlier. Consider a phishing campaign that uses hundreds of slightly different emails generated by AI. To a human analyst, each message may appear unique. An AI-powered system, however, can identify common patterns, infrastructure overlaps, or behavioral similarities that point to a shared source.

The same applies to malware analysis. Attackers may use AI to rapidly generate variants that evade signature-based detection, but defenders can use AI, including various machine learning models, to identify underlying similarities in code DNA, execution patterns, or command-and-control behavior. What appears to be dozens of unrelated incidents may, in fact, be a single campaign.

Likewise, AI should be capable of correlating weak signals across multiple telemetry sources—an unusual login pattern, a suspicious endpoint event, and seemingly benign cloud activity. Individually, these signals may not raise any alarms. Together, however, they can reveal a broader intrusion attempt and help analysts determine whether they are dealing with an opportunistic cybercriminal, a ransomware affiliate, or a more sophisticated threat actor.

This is close to how ESET PROTECT XDR correlates detections, thanks to its AI-powered detection engine and ESET LiveCortex.

The result is a new gray zone for cyber defense. But uncertainty does not automatically favor attackers. If anything, it raises the value of intelligence, correlation, and context—the very capabilities modern SOCs are investing in today.

Attribution and deterrence in the AI era

Cybersecurity practitioners have traditionally treated attribution as one of the end goals of intelligence. But in an AI-enabled environment, organizations may need to prioritize action over certainty. In other words, an SOC needs less of “this was done by group A,” and should instead focus on:

  • Is the campaign still active?
  • What infrastructure is being used?
  • What systems are typically targeted?
  • What should be blocked immediately?

Mind you, AI doesn’t run campaigns on its own; even though independent agentic action is not completely off the table, someone must still have turned it on. But, regardless of who is behind an attack, defenders often need to make decisions without waiting for high-confidence attribution.

In other words, it’s less about what an intelligence feed says about an attacker’s profile, and more about what’s happening now, combined with the visibility that XDR provides.

Consider a ransomware intrusion. An SOC may not immediately know which group is responsible, but it can still identify the command-and-control infrastructure, detect lateral movement attempts, isolate compromised hosts, and block malicious domains. The attacker may remain anonymous, yet the operation becomes slower, more expensive, and more likely to fail.

The same principle applies to AI-assisted phishing. Suppose an attacker generates thousands of personalized lures using generative AI. A defender does not need to know who wrote each email. By identifying common delivery infrastructure, shared behavioral indicators, or recurring targeting patterns, the SOC can disrupt the campaign before it reaches its objectives.

Deterring the gray zone

And this is where deterrence enters the picture. Unlike attackers, defenders control the environment being targeted. They can harden systems, improve visibility, and shorten dwell time, thereby increasing the operational costs of an intrusion. Even if attribution becomes less certain, organizations can still make themselves more difficult and less profitable targets, potentially discouraging attackers from continuing their efforts.

In that sense, attribution remains valuable, but not solely for assigning blame. Its practical value lies in helping defenders understand adversary behavior, anticipate future activity, and deploy countermeasures before an attack succeeds. 

The goal, in a sense, is to create enough friction so that attackers are forced to spend more time, consume more resources, and accept greater risk.

What SOCs should prioritize now

Now, AI is making attribution less certain, so SOCs must focus on extracting signal from noise and acting without waiting for certainty. In practice, this means:

  • Placing timely intelligence first: Real-time threat intelligence can help SOCs understand which techniques are gaining traction, what infrastructure is being abused, and which sectors are being targeted as of the last few minutes. Even when defenders cannot confidently identify a specific threat actor, they can still recognize patterns associated with campaigns or operational styles.
  • Investing in richer telemetry: Opt for clarity. Procure intelligence feeds aligned with your organization’s threat profile, whether that means a PUA feed that tracks malicious tool use or a botnet feed that enhances early-stage detection.
  • Focus on behavior, not indicators: IP addresses change and domains get abandoned. Attackers have always rotated technical indicators, but AI may accelerate this process dramatically. Instead of relying solely on static IoCs, SOCs should invest in detecting behaviors: unusual privilege escalation, abnormal authentication patterns, suspicious lateral movement, or unexpected data access. While attackers can generate endless variations of tools and infrastructure, changing operational behavior is often much harder.
  • Use AI to assist: Modern SOCs already face alert fatigue. If AI-assisted attacks increase the volume of events requiring investigation, analysts will need help clustering related incidents, identifying patterns across telemetry, and surfacing activity that merits human attention. This is useful, even for APT intelligence, for rapid insight into sophisticated attackers.

Conclusion

The future of attribution may be less about naming the attacker and more about understanding the attack well enough to stop it. AI will blur fingerprints, muddy old signals, and give attackers more room for plausible deniability. 

But defenders are not left guessing in the dark. If an SOC can detect a campaign earlier, contain it faster, and make the attacker work harder for every step forward, then intelligence is translating into a defensive advantage. Not perfectly, perhaps, but well enough to become a force multiplier for deterrence.

eset-threat-intelligence_banner

FAQ: AI and attribution

What is cyber attribution, and why does it matter?

Cyber attribution is the process of identifying who is behind a cyberattack. For SOCs, attribution matters because it helps shape response priorities, guide threat hunting, and inform longer-term security investments.

How does AI make cyber attribution harder?

AI can make attribution harder by blurring signals analysts have traditionally relied on, such as writing style in phishing emails, malware variations, and operator behavior. As attackers use generative AI and automation across the attack chain, campaigns may become harder to link confidently to a specific group or individual.

Can AI also help defenders improve attribution?

Yes. AI can help defenders correlate weak signals across large volumes of telemetry, cluster related incidents, detect behavioral similarities, and surface patterns that might be missed by human analysts alone. In this sense, AI can both complicate attribution and strengthen intelligence-led defense.

Should SOCs wait for perfect attribution before responding?

No. In the AI era, SOCs often need to act before attribution reaches high confidence. Blocking malicious infrastructure, isolating compromised hosts, detecting lateral movement, and disrupting active campaigns can reduce attacker success, even when the exact operator remains unknown.

What should SOCs prioritize as AI changes the threat landscape?

SOCs should prioritize timely threat intelligence, richer telemetry, behavior-based detection, and AI-assisted investigation. These capabilities help teams extract signal from noise, respond faster, and raise the operational cost for attackers, even when attribution becomes less certain.