Four of Five AI Security Numbers Didnt Survive a Trace

   / @thehambergerreport   Four of five headline AI agent security statistics circulating this year do not survive being traced to source. OWASP's 2026 LLM Top 10 admits its own prompt injection ranking would fall out of the top ten entirely on incident data alone, without the community vote supplying 75 per cent of its score. The widely repeated claim that autonomous AI agents outnumber humans 82 to 1 traces to a company that had already agreed to acquire the company whose figure it cited, four months before publishing it, undisclosed. A Recorded Future report cited for a 340 per cent surge does not appear to exist. One benchmark, AIShellJack, names its method and its 314-payload sample, and survives. A working proof-of-concept in verification-first AI governance: every episode is produced through gated control points that test real-time verification, source-checking, and containment models, with a person confirming each gate. Practical analysis by Andreas Hamberger, enterprise architect, technology strategist, and author of four books including Lethal By Design. Lethal By Design: https://hamberger.short.gy/lethalbyde... 🌐 thehambergerreport.com 💼 linkedin.com/in/andyhamberger Chapters: 00:00 Cold open 00:24 Introduction 02:02 The ratio that outlived its source 03:26 Two surveys and a falling number 04:52 The one number that survived 06:18 What a public repository proves 07:42 The framework that demands proof 09:17 Before it goes in your deck 10:33 Wrap-up [1] Help Net Security. "OWASP 2026 LLM Top 10 released." 6 August 2026. https://www.helpnetsecurity.com/2026/... [2] Palo Alto Networks. "2026 Predictions for Autonomous AI." 25 November 2025. https://www.paloaltonetworks.com/blog... [3] PR Newswire. "Palo Alto Networks Forecasts 6 Predictions on Securing the New AI Economy for 2026." 18 November 2025. https://www.prnewswire.com/news-relea... [4] CyberArk. "Machine Identities Outnumber Humans by More Than 80-to-1, New Report Exposes the Exponential Threats of Fragmented Identity Security." 23 April 2025. https://www.cyberark.com/press/machin... (this URL now redirects to Palo Alto Networks' Idira identity-security platform page, https://www.paloaltonetworks.com/idira, which cites the successor 109-to-1 figure) [5] Gravitee. "State of AI Agent Security 2026 Report: When Adoption Outpaces Control." 4 February 2026. https://www.gravitee.io/blog/state-of... [6] Gravitee. "State of AI Agent Security." Updated April 2026. https://www.gravitee.io/state-of-ai-a... [7] Token Security. "65 Percent of Enterprises Have Already Experienced AI Agent Security Incidents." 21 April 2026. https://www.token.security/blog/65-pe... [8] The Paypers. "CSA survey finds 82% of enterprises have unknown AI agents in their environments." 28 April 2026. https://thepaypers.com/fraud-and-finc... [9] Center for Internet Security. "Prompt Injections: The Inherent Threat to Generative AI." 18 March 2026. https://www.cisecurity.org/insights/w... (checked directly and confirmed to contain no percentage or growth figures) [10] Recorded Future. "Introducing the 2025 State of Threat Intelligence Report." https://www.recordedfuture.com/blog/i... (checked directly and confirmed to contain no mention of prompt injection or the figures commonly attributed to it) [11] Liu, Zhao, Lyu, Zhang, Wang and Lo. "'Your AI, My Shell': Demystifying Prompt Injection Attacks on Agentic AI Coding Editors." arXiv preprint 2509.22040, first posted 26 September 2025, updated 28 April 2026. https://arxiv.org/abs/2509.22040 [12] National Cyber Security Centre New Zealand. "Careful Adoption of Agentic AI Services." 1 May 2026. https://www.ncsc.govt.nz/protect-your... [13] Kordia. "Biggest AI cyber threat may be coming from inside your business, new report finds." 9 March 2026. https://www.kordia.co.nz/news-and-vie... [14] UK Ministry of Defence. Joint Service Publication 936, Dependable Artificial Intelligence in Defence. (No specific edition date or URL captured in this episode's source material; cited by title and issuing authority only, per URL integrity practice, rather than an invented or reconstructed link.)