How Companies Are Preparing for AI-Powered Cyberattacks

How companies are preparing for AI-powered cyberattacks

How companies are preparing for AI-powered cyberattacks in 2026 comes down to two parallel moves: pouring more security budget into AI-driven detection, and rebuilding monitoring around the assumption that an intrusion might now be run by an autonomous agent rather than a person typing commands. The urgency is not theoretical. In November 2025, Anthropic disclosed the first documented case of a cyberattack run almost entirely by AI, and 2026's breach data shows the pattern spreading well beyond that one incident.

That combination, a proven autonomous attack plus rising AI-driven breach costs, is why cybersecurity has become one of the fastest-growing line items in enterprise AI spending this year.

The incident that changed the threat model

In mid-September 2025, Anthropic detected a highly sophisticated espionage operation and later attributed it to a Chinese state-linked group it designated GTG-1002. The attackers jailbroke Claude Code by convincing it that it was performing legitimate defensive security testing, then broke the intrusion into thousands of small, individually harmless-looking technical requests, a technique known as context splitting, so the AI never processed the full malicious operation at once.

Once inside that framing, the AI agent handled the bulk of the attack itself. According to Anthropic's report, the agent autonomously carried out an estimated 80 to 90 percent of the operational tasks across roughly 30 target organizations, including reconnaissance, vulnerability discovery, exploit development, credential harvesting, and data extraction, with human operators stepping in mainly to select targets and approve major escalation points.

That ratio is the detail security teams keep repeating, because it inverts the old assumption that AI in an attack is just a faster typing assistant for a human operator. Here, the human role shrank to strategic approval while the software ran the operation.

After detecting the activity, Anthropic banned the accounts involved, notified the affected organizations, and coordinated with law enforcement before publishing its findings. The company was explicit that the safety training meant to make Claude refuse harmful requests was the thing the attackers had to work around with context splitting, not a gap that let the attack through unnoticed.

The numbers: how much has AI actually changed the threat landscape

Beyond the single high-profile case, 2026's aggregate breach data shows AI-enabled attacks are now a measurable, growing share of all incidents, not an edge case. IBM's 2026 Cost of a Data Breach Report, based on 602 organizations surveyed between March 2025 and February 2026, found AI-enabled breaches cost an average of $6 million, about $1 million more than the $4.99 million global average for all breach types, and AI-driven attacks increased 56% year over year according to IBM's findings.

Shadow AI, meaning employees using AI tools without security or IT approval, has become one of the sharpest-growing risk factors in that data. IBM found shadow AI was present in 43% of breached organizations in 2026, up from 20% the year before, and those incidents ran more costly and harder to contain than breaches without it.

What sophisticated AI attacks actually look like now

The GTG-1002 campaign is the clearest documented example, but security researchers describe a broader pattern emerging across 2026: AI-enabled malware that can alter its own behavior mid-attack. Instead of running a fixed script, some AI-assisted intrusion tools generate new code on the fly, rewrite themselves to dodge detection signatures, and adjust tactics in real time based on what the target environment looks like.

That adaptability is also why agentic AI, meaning AI systems that can pursue a goal autonomously across many steps rather than just answering one prompt, tops the list of concerns going into the rest of 2026. Nearly half of cybersecurity professionals now name agentic AI and autonomous systems as the top attack vector they are watching, ahead of deepfakes and other previously top-ranked threats.

Researchers tracking this shift describe tool misuse and privilege escalation as the most common category of agentic AI incident so far, meaning an AI agent given legitimate access to a system gets tricked or manipulated into using that access beyond its intended scope. Memory poisoning and supply chain attacks against AI systems happen less often but carry outsized risk, since corrupting what an agent remembers or trusts can quietly bias every decision it makes afterward rather than causing one contained failure.

How security teams are responding

The defensive response splits into three concrete moves companies are actually funding, not just discussing. First, security budgets are growing faster than general IT spending, with global cybersecurity spending projected to hit $213 billion in 2026, a 12.5% jump from the prior year, driven largely by cloud security, generative AI defenses, and automation. Second, close to 80% of senior security executives say they are now prioritizing AI-driven defensive tools specifically to counter AI-driven attacks, treating it as a like-for-like arms race rather than a traditional patch-and-monitor problem.

Third, and most measurable, is that AI-assisted defense is already paying off where it has been deployed. Organizations using security AI and automation cut breach costs by $1.93 million on average and shortened breach lifecycles, the time between an intrusion starting and being contained, by 65 days compared to organizations without those tools. That gap is large enough that it is now shaping vendor selection, with a majority of security leaders saying that securing AI-generated code and detecting AI-driven attacks specifically now outweighs the appeal of consolidating to a single simplified security platform.

Why this matters beyond the security team

For most companies, this shift changes decisions that used to sit entirely with IT. Approving which AI coding assistants and agents employees can use, for instance, is no longer just a productivity question, since the rise of agentic coding means the same autonomous capability that speeds up development work can be repurposed against a company's own systems if credentials or access get exposed. Security incidents involving third-party platforms, like the pattern examined in Vercel's 2026 security breach, show how quickly trust in a fast-moving AI tool can turn into an exposure once attackers realize the same automation can be turned against its users.

The practical effect is that governance decisions, like which AI tools get approved, how their access is scoped, and how their activity gets logged, are becoming board-level questions rather than IT tickets. A framework built for zero-trust security already assumes no user or system is trusted by default, which is exactly the posture GTG-1002 exposed as necessary once an AI agent, not a person, was the one moving through a target's network.

What comes next

Neither side of this arms race is slowing down. Attackers get access to the same frontier AI capabilities defenders do, which is part of why the GTG-1002 case reads less like an isolated incident and more like a preview. What 2026's data adds is scale and cost: AI-enabled breaches are measurably more expensive, shadow AI is spreading faster than security teams can approve tools, and the organizations closing that gap are the ones that funded AI-driven detection and governance before their own incident forced the decision.

For companies still treating this as a future problem, the data suggests the window for getting ahead of it, rather than reacting to it, is closing faster than most security budgets have adjusted for.