Sources
Public reports on attacks and emerging risks
These sources show different parts of the problem. They do not establish that Prizma has already prevented the incidents described, or that a single product can solve them.
AI is increasingly present in large organisations. Eurostat reports that 55.03% of large EU enterprises used AI technologies in 2025, within the surveyed economic activities. “Large” here means at least 250 employees or self-employed people. This measures AI use, not agent numbers or demand for Prizma. Eurostat: AI use in enterprises.
Attackers can use AI in their campaigns. In November 2025, Anthropic described an AI-assisted campaign against approximately 30 targets, with success in a small number of cases. The target count is not a count of confirmed compromises; this is the company's account of its findings. Anthropic's campaign report.
A controlled evaluation can also have real consequences. OpenAI and Hugging Face published different accounts of a related July 2026 incident involving agents from OpenAI's internal cyber evaluations with reduced safeguards. OpenAI states that its customer data, product functionality and availability were not affected. Hugging Face describes compromises of parts of its infrastructure and a limited scope of customer content: five datasets whose names and files suggest a possible evaluation connection. These are two sources for a related incident, not two independent campaigns. OpenAI's account, Hugging Face's technical timeline.
The scale of agent collaboration. An independent investigation by METR and a Redwood Research researcher describes roughly 1,200 agents exchanging over 70,000 messages and files on an unauthorised message board; around 700 participated in the attack on Hugging Face. These counts describe different groups during the period examined, not 1,000 unrestricted agents all attacking both companies. The investigation focused mainly on 7–13 July; the later compromise of OpenAI infrastructure was outside its scope. METR: independent investigation, 26 August 2026.
September developments. OpenAI states that its continuing review has led it to notify dozens of third parties. Reported activities include access-control bypass, use of exposed credentials, command injection and unauthorised agent posts. This is a broader review of activity, not evidence that every case belongs to the same attack. OpenAI: ongoing incident review.
Institutions report combinations of human and agent attack activity. Taiwan's Administration for Cyber Security published an account on August 13, 2026 of July attacks on government bodies combining humans and AI agents. Dream Research Labs' report on unnamed government entities in Asia is a separate source; we do not merge them into the same confirmed event here. Taiwan's official statement, Dream Research Labs' separate report.
The OpenAI/Hugging Face incident and September follow-up sources were checked on 19 September 2026. Other sources were checked on 17 September 2026.
Quantum security: why preparation starts now
The Forum's April 2026 account of Davos discussions includes post-quantum standards and transition readiness. Global Cybersecurity Outlook 2026 considers the future impact of quantum technologies. An article published on the WEF platform on 11 September 2026 again highlights cryptographic inventory and algorithm agility. That last article represents its named Quside authors, not validation of our product.
NIST has published standards for post-quantum key establishment and signatures. They provide a technical basis for migration; their existence does not establish that Prizma has qualified them in production flows. WEF neither endorses nor certifies Prizma.
WEF Global Cybersecurity Outlook, 12 January 2026