Sophos’s AI Security 2026 report, drawn from observations across more than 625,000 customer organizations, makes one point clear: artificial intelligence is speeding up cybercrime, not reinventing it. It is changing the clock of cybersecurity, not its fundamentals. Attackers still need to break in, move around, and steal data. They are simply doing it faster.
The case that set the tone
The report opens with a striking incident from May 2026. Sophos found an attacker running what amounted to a software factory inside a customer’s network. Around a dozen AI agents, coordinated through a commercial coding assistant, wrote and tested malware against endpoint protection from Sophos, CrowdStrike and Windows Defender. They produced nearly 80 modules and more than 70 evasion techniques. The same operator later used the output to deploy ransomware and steal data.
The tradecraft was familiar. The speed was not: weeks of manual work shrank into days. Sophos also noted that the evidence did not fully support the agents’ claims of near-universal success, reinforcing the report’s point that AI speeds execution more than it transforms the attack itself.
Criminals are adopting AI, unevenly
On underground forums, threat actors are buying brokered API keys for ChatGPT, Claude and Grok, trading jailbreak tips, and even hiring prompt engineers. Ads for AI voice bots and fake AI-generated dating personas show social engineering has become operational. Yet some criminals remain skeptical, worried AI will cut their own job opportunities, underscoring uneven adoption rather than a total shift in cybercrime.
The report is careful on one point: fully autonomous AI attack campaigns remain unconfirmed at scale. Even the most cited case, an Anthropic-documented campaign automating most of a data theft operation, is described as anomalous rather than a trend, which keeps the report focused on acceleration, not reinvention.
The enterprise is the softer target
The most practical finding is that AI tools themselves are being attacked. Compromised developer extensions, poisoned software packages, and fake AI download pages are already hitting organizations. The Salesloft/Drift and Vercel incidents showed how a chatbot’s access tokens can open the door to core business systems.
The governance gap is wide. According to research cited in the report, 75% of security leaders had found shadow AI tools, only 1% of enterprises have a dedicated AI security budget, and just 16% of large firms effectively govern the access their AI agents have, despite 71% having deployed them.
Patching windows are shrinking
Exploitation is outrunning repair. One Ivanti vulnerability was exploited within 24 hours of a public proof of concept, and a LiteLLM flaw within 36 hours. The report says CISA’s three-day patching deadline should be viewed as a floor, not a target.
Defenders get faster too
AI reasoning models can compress hours of analyst investigation into minutes. But they are not a cure-all. They lack memory of past incidents, behave more like junior analysts, and sit on top of, not in place of, layered behavioral detection. Sophos describes a pipeline that reduced 11.8 trillion events to about 81,500 high-priority alerts. Another bottleneck is human: investigation may take minutes, but change tickets and maintenance windows can still delay the response by days.
What Leaders should do?
Sophos recommends starting with visibility: inventory every AI tool, what it can access and which credentials it holds. It also proposes seven controls for agentic AI, including sandboxing, credential isolation, egress monitoring and human-gated approval for irreversible actions. For any deployment, ask a simple question: what is the worst credible outcome, and can it be reversed?
The takeaway
The report cites the UK AI Security Institute’s findings of a roughly sixfold increase in autonomous offensive capability over 18 months, so complacency is not an option. Sophos’s point is clear: adjust your tempo now, or fall further behind than you assume.
One note for readers: as a vendor report, its findings reflect Sophos’s own vantage point and commercial interests, though it is notably restrained about hype.
