AI
From a LinkedIn Post to Industry Standard
Steve Wilson expected 10 people when he posted about LLM security on LinkedIn. He got 200 on a Slack channel within seven days. The OWASP LLM Top 10 project now has over 3,000 LinkedIn followers, and Wilson estimates more than 100,000 developers and AppSec professionals have read the document. A share-of-voice analysis found 58% of articles about securing LLMs reference the Top 10. OWASP joined the NIST AI Safety group, and MITRE ATLAS added entries that map directly from the list.
The Vulnerabilities Companies Keep Shipping
The industry is rushing to jam chatbots into software without understanding the risks. Microsoft Copilot and Slack both fell to indirect prompt injection. Wilson marks the shift: it moved from small companies making embarrassing mistakes to the most sophisticated software companies making the same ones. The AI supply chain amplifies the problem. Studies now report tens of thousands of poisoned models on Hugging Face. Wilson predicts Apple Intelligence will hit these same issues, given no public signals the company has addressed them.
How the Project Grew Beyond the Top 10
The project now runs four initiatives alongside the list. A data collection initiative builds proof points and maps the Top 10 to industry standards. A research initiative led by Rachel James runs joint work with the University of Illinois to fill evidence gaps. A third produces guidance for building AI security centers of excellence. A fourth, approved three weeks before this talk, covers red teaming. One observation: AI builders treat red teaming as a novel AI concept because it appeared in President Biden’s executive order. AppSec teams already hold the foundations.
What Version 2.0 Changes
The 2024 edition targets a November release. Interim voting shows where risks have moved. Supply chain jumped to second, right behind prompt injection. It was an abstract worry 12 months ago and is now top of mind. Excessive agency is rising fast. Startups dedicated entire conferences to AI agents, and Wilson argues winning those deployments depends on doing them securely, not on feature count. The developer participation gap persists: more CISOs than software developers are engaged with the project.
AI Regulations Come Down to Three Fears
AI regulations are arriving across multiple jurisdictions. Wilson distills them to three concerns. First: data use, covering personal information and unauthorized web scraping. Second: job displacement, including the EU AI Act’s ban on AI-based hiring decisions. Third: safety, from AI in medical diagnosis to unconstrained large-scale model training. California bills set parameter count thresholds that trigger mandatory government approval. Deep fakes compound the threat. A CISO at Wilson’s company ran an interview where the candidate was wearing a deep fake, possibly a North Korean agent.
Q&A
Are there mappings from the Top 10 to existing controls and verification tools? The data collection initiative has controls mapped to CVSS and MITRE standards; the upcoming solutions guide will list tools by category, including supply chain. ▶ Watch (34:23)
When will version 2.0 be released? Wilson expects November 2024, after outstanding volunteer homework is submitted and a final voting round narrows the list. ▶ Watch (36:50)
Will the project produce an AI-specific secure development lifecycle? Chapter 12 of Wilson’s book covers a Responsible AI Software Engineering Framework; the solutions guide also maps LLM SecOps steps to the development process. ▶ Watch (43:51)
Notable Quotes
not on my bingo card 12 months ago Steve Wilson · ▶ Watch (8:54)
software supply chain is a dumpster fire Steve Wilson · ▶ Watch (15:48)
developers don’t care um Steve Wilson · ▶ Watch (23:56)
Key Takeaways
- The OWASP LLM Top 10 reached 58% share of voice in LLM security writing within one year of launch.
- Supply chain risk jumped to second in version 2.0 voting; excessive agency is rising with the AI agent boom.
- Four new project initiatives cover data collection, research, red teaming, and an AI security center of excellence.
About the Speaker(s)
Steve Wilson is the founder and project leader of the OWASP LLM Top 10, where he assembled over 1,000 experts to build the leading security reference for generative AI. He is Chief Product Officer at Exabeam and published “The Developer’s Playbook for Large Language Model Security” through O’Reilly.