Audited cloud security, but exploited flaws and AI rules that cannot be viewed
Overall rating “limited”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
Arista documents general risk oversight by the Audit Committee, but its AI policy and AI committee are only company statements, and figures and external assurance are missing.
What is documented
- The Audit Committee has a mandate for cybersecurity, information security, data protection and environmental and social matters; the board oversees risk management1 2
- According to the company, an internal AI policy exists with guidelines on security, fairness, intellectual property and human oversight2 3
- According to the proxy summary, there is an AI governance committee of management, without details of its composition or reporting line2
- A reporting portal and an anonymous hotline in 32 languages are available; retaliation is prohibited6 3
- Employees receive training on data protection and cybersecurity as well as AI workshops3
The case against
Company statements without evidence
Sources for Responsibility (6)
- Arista Networks, Inc. (SEC EDGAR): Definitive Proxy Statement (DEF 14A) 2026 · 16 April 2026 · primary source
- Arista Networks, Inc. (SEC EDGAR): Annual Report to Security Holders (Form ARS) FY2025 incl. proxy summary · 3 November 2026 · secondary source
- Arista Networks, Inc.: 2025 Corporate Responsibility Report · 3 November 2026 · secondary source
- Arista Networks, Inc. (SEC EDGAR): Form 10-K for the fiscal year ended December 31, 2025 · 17 February 2026 · primary source
- Arista Networks, Inc.: Arista Trust Center · 3 November 2026 · secondary source
- Arista Networks, Inc.: Code of Ethics and Business Conduct (as of 2025-10-23) · 23 October 2025 · secondary source
2 · Safety
Arista documents a vulnerability response team, SOC 2 Type 2 and product certifications, but four flaws were actively exploited in 2026, and AI-specific testing of AVA is missing.
What is documented
- CloudVision as-a-Service has SOC 2 Type 2 with an annual external audit, encryption, tenant separation and external penetration tests4
- The PSIRT is the central reporting point with PGP, credits researchers and assigns CVEs in a coordinated way12
- Advisories contain indicators, attacker IP addresses and guidance on preserving evidence8 9
- Network products hold FIPS 140-3 certificates and DoDIN listings, without an AI link3
- According to the product brief, AVA is designed for decision support with human oversight and guardrails1 2
The case against
Actively exploited vulnerabilities
AI-specific safeguards not documented
- Testing, validation, red-teaming and monitoring of AVA and the purchased language models are not documented
- There is no information on the protection of models, prompts and the MCP interface of AVA1
Sources for Safety (14)
- Arista Networks, Inc.: Arista AVA Product Brief · 3 November 2026 · secondary source
- SiliconANGLE: Arista expands AVA for AI Ops use cases · 10 December 2025 · secondary source
- Arista Networks, Inc.: Arista Federal – Security Compliance · 3 November 2026 · primary source
- Arista Networks, Inc.: CloudVision as-a-Service Security Whitepaper · 5 March 2026 · primary source
- Arista Networks, Inc.: Arista Trust Center · 3 November 2026 · secondary source
- board-cybersecurity.com: Arista Networks, Inc. 10-K Item 1C. Cybersecurity – 2026-02-13 · 13 February 2026 · secondary source
- Arista Networks, Inc. (SEC EDGAR): Definitive Proxy Statement (DEF 14A) 2026 · 16 April 2026 · primary source
- Arista Networks, Inc.: Security Advisory 0144 (VeloCloud Orchestrator, CVE-2026-16812) · 27 July 2026 · primary source
- Arista Networks, Inc.: Security Advisory 0183 (VeloCloud Orchestrator, CVE-2026-93952) · 22 September 2026 · primary source
- The Hacker News: Arista warns of critical VeloCloud SQL injection flaw exploited as zero-day · 22 October 2026 · secondary source
- Eclypsium: Arista EOS tunnel decapsulation flaw: no patch · 16 June 2026 · secondary source
- Arista Networks, Inc.: Security Advisories (overview page, PSIRT) · 9 September 2026 · primary source
- Security Affairs: U.S. CISA adds Arista VeloCloud Orchestrator and Fortinet FortiOS flaws to its Known Exploited Vulnerabilities catalog · 28 July 2026 · secondary source
- The Hacker News: CISA adds Cisco, Chrome and Arista flaws to KEV catalog · 9 June 2026 · secondary source
3 · Transparency
Arista names AI risks and internal areas of use in general terms, but the language models in AVA remain unnamed, and the privacy policy does not mention AI.
What is documented
- The 10-K discloses AI risks in general terms1
- The Corporate Responsibility Report names the internal areas of AI use, including the pre-screening of applicants2
- AVA shows its reasoning steps; according to the brief, conclusions are only as good as the telemetry data4
- CloudVision as-a-Service uses anonymised, aggregated data with machine learning for anomaly detection5
The case against
Missing model information
- AVA uses purchased, unnamed language models; providers and data flows are not disclosed4
- No system card or documented limitations and error rates of AVA were found
No labelling
- The privacy policy does not mention AI; no notices to applicants about AI-supported pre-screening were found3
Sources for Transparency (5)
- Arista Networks, Inc. (SEC EDGAR): Form 10-K for the fiscal year ended December 31, 2025 · 17 February 2026 · primary source
- Arista Networks, Inc.: 2025 Corporate Responsibility Report · 3 November 2026 · secondary source
- Arista Networks, Inc.: Privacy Policy (Last Updated July 2026) · 3 November 2026 · secondary source
- Arista Networks, Inc.: Arista AVA Product Brief · 3 November 2026 · secondary source
- Arista Networks, Inc.: CloudVision as-a-Service Security Whitepaper · 5 March 2026 · primary source
4 · Why the harm rule does not apply
The harm rule does not apply: there is no judgment, authority decision or fine concerning harm from AI, and the exploited flaws have no documented link to AI.
5 · Controversies (2)
- mediumBetween July and October 2026 Arista reported three actively exploited critical vulnerabilities in the acquired VeloCloud Orchestrator; CISA added one to its KEV catalogue. · 27 July 2026 · open
- mediumArista reported the exploited EOS vulnerability CVE-2026-7473 and announced it would not provide a software fix; CISA added it to its KEV catalogue. · 5 May 2026 · open
6 · What remains open
- Board responsibility for AI and composition of the AI governance committee
- Content of the internal AI policy, risk classification and approval procedures
- Testing, red-teaming and protection of AVA and the names of the language models used
- Rules and transparency on AI-supported pre-screening of applicants
- Items 1C and 3 of the 10-K in the original and the number of customers affected by the exploited flaws
7 · How it was assessed
Cut-off date: 3 November 2026 · 43 searches, 31 fetches.
Primary documents read (9)
- Form 10-K FY2025 (cover page, Item 1, Item 1A, Human Capital; three retrievals) and filing index
- Definitive Proxy Statement 2026 (three retrievals)
- Form ARS FY2025 incl. proxy summary (three retrievals)
- 2025 Corporate Responsibility Report (three retrievals)
- Form 10-Q Q2 2026 (to Item 2, two retrievals)
- Form 8-K voting results AGM 2024, 2025, 2026
- Code of Ethics and Business Conduct (2025-10-23), Privacy Policy (July 2026), Trust Center
- AVA Product Brief, CloudVision-as-a-Service Security Whitepaper (2026-03-05), Federal Security Compliance
- Security Advisories 0144 and 0183, PSIRT overview page
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0