Sophos has published its AI Security 2026 report in which it reveals that cybercriminals are using artificial intelligence (AI) to reduce attack timelines from weeks to days.
The report reveals that the most immediate impact of AI on cybercrime is speed, as well as an increase in attacks that use identity as the main initial access vector, rather than inventing new attack types at this stage.
“Attackers still need initial access, they still move laterally, and they still exfiltrate data through observable channels. What has changed is the timing of this,” said John Peterson, CTO at Sophos. “For the first time, we saw AI being actively used as an operational capability multiplier. While the tools and techniques were already familiar, the speed of development, testing and iteration was substantially different. That’s the AI threat that security teams must prepare for. It involves faster cycles and shorter response times, which puts more pressure on defenders to detect and contain activity before impact occurs.”
Key takeaways from the report
- AI is reducing attack timelines and accelerating operational readiness.
- Enterprise AI identities, OAuth tokens, agents, APIs, and development tools are becoming high-value targets.
- AI-assisted social engineering and deepfakes are already operational tools.
- Malicious actors are incorporating AI into underground markets, recruitment, prompt engineering, jailbreaking, malware development workflows, and criminal services.
- AI development infrastructure and supply chains are in the spotlight.
AI in the hands of attackers
In line with one of the most significant findings of the report, Sophos discovered a campaign identified as STAC6994 that was actively using AI to power its operations; This is one of the first reliable demonstrations of what is happening.
The malicious actor was running a software development operation within a Sophos customer network and used approximately 12 AI agents to write and test attacks against Sophos endpoint agents CrowdStrike and Microsoft Defender.
They produced almost 80 modules and more than 70 evasion techniques, reducing to a few days what would have taken a human being weeks. This dramatically accelerated the time required for attack techniques to reach operational maturity.
However, that intelligence gathering effort allowed Sophos to get ahead of the threat and thwart attacks before they spread into the real environment.
AI identities, a new attack surface
The report notes that AI adoption in businesses is the fastest growing source of new risks. As scheduling agents and AI assistants and models gain privileged access to core systems, attackers focus on the trust, credentials, and access permissions associated with those systems. As a result, AI identities, agents, OAuth connections, and API keys are increasingly becoming a high-value attack surface, and their control is not keeping pace.
Attackers are thus creating new avenues of access to enterprise networks by compromising OAuth tokens, AI service credentials, development tools, and exposed AI infrastructure. This shows that AI is now as much an identity, governance and supply chain issue as it is a model safety issue.
This conclusion is also reflected in the recent Sophos State of Ransomware 2026 report, which reveals how, for the first time in more than three years, identity has become the main attack vector.
Social engineering powered by AI and deepfakes
Likewise, AI-assisted social engineering and deepfakes are making scams more scalable, more convincing in different languages, and significantly cheaper to generate. Incidents highlighted in the report include an AI-related investment scam that lured a UK-based victim to a fake AI-powered investment platform after months of AI-generated conversations and coordinated messages. The victim ended up losing hundreds of thousands of pounds.
AI development infrastructure is also being directly targeted by attacks involving compromised development tools and credential-stealing malware. Supply chain risks related to models, training data provenance, MCP servers, and inference infrastructure are also becoming more prevalent.
AI-assisted social engineering and deepfakes are making scams more scalable
“The report makes clear that AI security is no longer limited to the behavior of models or future speculative risks. AI is being actively integrated into criminal workflows and social engineering operations, as well as into enterprise software development and the identity systems of legitimate organizations. This means that the threat is here and now,” concludes the head of Sophos. “As cutting-edge models continue to advance, the coming months will be determined by how quickly organizations can regulate the use of AI, protect identities and connections related to it, and keep pace with attackers who are able to adopt new capabilities very quickly.”
The report is based on findings from cases managed by Sophos
