AI oversight at board level and ISO 42001, but repeated data leaks via AI agents
Overall rating “solid”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
The Cybersecurity & Privacy Committee has primary responsibility for AI, and a Trusted AI Review with ISO 42001 certification documents implementation; externally audited AI disclosures and complaint channels for affected people are missing.
What is documented
- According to the proxy, the Cybersecurity & Privacy Committee has primary responsibility for AI policies and programmes and met eight times in FY 20261
- An Office of Ethical and Humane Use with its own head has existed since 2019, and an external Ethical Use Advisory Council since 20182
- In the Trusted AI Review, ethical risks are treated as bugs with severity levels and can stop a launch2
- In FY 2026 more than 370 internal and more than 240 customer-facing AI applications were reviewed3 4
- The ISO 42001 certification is valid until 2028 and includes annual surveillance audits7 8
The case against
Missing external accountability
Sources for Responsibility (11)
- Salesforce, Inc. / U.S. SEC: Proxy Statement (DEF 14A) 2026 · 3 November 2026 · primary source
- Salesforce, Inc.: Trusted AI and Agents Impact Report (first edition) · 1 February 2025 · primary source
- Salesforce, Inc.: Blog: Second Annual Trusted AI & Agents Impact Report · 11 May 2026 · secondary source
- Salesforce, Inc.: FY26 Stakeholder Impact Report · 31 January 2026 · primary source
- Wind River (licence list, copy of Salesforce policy): Salesforce Artificial Intelligence Acceptable Use Policy (Last Updated: August 5, 2025) · 5 August 2025 · secondary source
- U.S. District Court N.D. Cal.: In re Salesforce LLM Copyright Infringement Litigation, Master File No. 3:25-cv-08862-CRB – Joint Case Management Statement and Order (ECF 45) · 16 March 2026 · primary source
- Salesforce, Inc. (Brad Arkin, Chief Trust Officer): Blog: Salesforce achieves ISO 42001 certification for Agentforce, Einstein and Slack AI · 8 October 2025 · primary source
- Salesforce Compliance Portal: ISO 42001:2023 – Salesforce AI Platform, Agentforce and Slack AI · 1 October 2026 · primary source
- Salesforce, Inc. (Kathy Baxter): Blog: Meet Salesforce's Trusted AI Principles · 28 April 2023 · secondary source
- Joseph Saveri Law Firm: Salesforce Large Language Model Litigation – Case Background · 15 October 2025 · secondary source
- Let's Data Science: Salesforce employees urge CEO to drop ICE · 13 February 2026 · secondary source
2 · Safety
External attestations, penetration tests and AI-specific testing are documented, but within two years researchers found three prompt injection flaws with data leakage, which were fixed.
What is documented
- The 10-K describes a NIST/ISO-based programme with penetration testing, external attestations and quarterly oversight by a board committee5
- Manual adversarial testing, fuzzing and external testing for two products are described for AI products1 2
- The Trust Layer provides masking, toxicity detection, an audit trail and zero data retention at LLM providers3
- All three known prompt injection flaws were fixed, and no exploitation was reported7 8 9
- After the 2025 attack campaign, Salesforce revoked tokens, blocked apps and restricted connected apps13 15
The case against
Flaws in its own AI products
- ForcedLeak (CVSS 9.4): Agentforce leaked CRM data to an expired domain via manipulated web-to-lead entries7
- SalesBleed: zero-click prompt injection leaked lead and account data via DNS queries, fixed in August 20268
- In 2024 Slack AI could be made to disclose data from private channels via indirect prompt injection9
Sources for Safety (18)
- Salesforce, Inc.: Trusted AI and Agents Impact Report (first edition) · 1 February 2025 · primary source
- Salesforce, Inc.: FY26 Stakeholder Impact Report · 31 January 2026 · primary source
- Salesforce Developers: Agentforce Developer Guide: Trust Layer · 3 November 2026 · primary source
- The Decoder: Salesforce executives signal declining trust in large language models · 25 December 2025 · secondary source
- Salesforce, Inc. / U.S. SEC: Form 10-K FY 2026, XBRL R26: Cybersecurity Risk Management and Strategy Disclosure · 31 January 2026 · primary source
- Salesforce Compliance Portal: ISO 42001:2023 – Salesforce AI Platform, Agentforce and Slack AI · 1 October 2026 · primary source
- The Hacker News: Salesforce Patches Critical ForcedLeak Bug Exposing CRM Data via AI Prompt Injection · 25 September 2025 · primary source
- Infosecurity Magazine: Zero-Click Vulnerabilities in Salesforce Agentforce Expose Wider AI Agent Risk · 25 September 2026 · primary source
- The Register: Slack AI can be tricked into leaking data from private channels · 21 August 2024 · primary source
- The Register: Fake IT support calls hit 20 orgs, end in stolen Salesforce data and extortion, Google warns · 4 June 2025 · primary source
- Krebs on Security: ShinyHunters Wage Broad Corporate Extortion Spree · 7 October 2025 · primary source
- United States Judicial Panel on Multidistrict Litigation: MDL No. 3164 (In re Salesforce, Inc., Customer Data Security Breach Litigation) and MDL No. 3170 – Transfer Order and Order Denying Transfer · 16 December 2025 · primary source
- BleepingComputer: Salesforce investigates customer data theft via Gainsight breach · 20 November 2025 · primary source
- Wind River (licence list, copy of Salesforce policy): Salesforce Artificial Intelligence Acceptable Use Policy (Last Updated: August 5, 2025) · 5 August 2025 · secondary source
- Salesforce Ben: Salesforce hardens connected apps security amid social engineering attacks · 21 August 2025 · secondary source
- Salesforce, Inc.: Blog: Protect Against Social Engineering · 12 March 2025 · secondary source
- The Register: Salesforce faces class actions over data theft · 26 September 2025 · primary source
- California Courts Newsroom (Daily Journal summary): Judge consolidates Salesforce data breach class actions in Northern California · 12 January 2026 · secondary source
3 · Transparency
Salesforce publishes model cards for its own models and documents limitations such as hallucinations, but discloses the training data of the xGen models and the third-party models it uses only incompletely.
What is documented
- The compliance portal contains 21 model cards for its own models, some updated in September 20266 7
- Research models such as xLAM have model cards with training dataset, licence and ethical notes8
- The Trust Layer documentation names hallucinations and bias as limitations and recommends human review4
- The AI acceptable use policy requires customers to disclose AI interactions and not to present AI content as human1
The case against
Gaps on data and models
- The training data of the xGen models is not disclosed and is the subject of a pending lawsuit9
- No system card for Agentforce and no transparency report with incident figures were found
Earlier weakness in disclosure
- In 2024 Slack's privacy principles provided for training global models on customer data by default, with opt-out only by email3
Sources for Transparency (9)
- Wind River (licence list, copy of Salesforce policy): Salesforce Artificial Intelligence Acceptable Use Policy (Last Updated: August 5, 2025) · 5 August 2025 · secondary source
- Salesforce, Inc. / U.S. SEC: Form 10-K for the fiscal year ended 2026-01-31 · 31 January 2026 · primary source
- TechCrunch: Slack under attack over sneaky AI training policy · 17 May 2024 · primary source
- Salesforce Developers: Agentforce Developer Guide: Trust Layer · 3 November 2026 · primary source
- Salesforce, Inc. (Kathy Baxter): Blog: Meet Salesforce's Trusted AI Principles · 28 April 2023 · secondary source
- Salesforce, Inc.: Blog: Model Cards for AI Model Transparency · 29 September 2020 · primary source
- Salesforce Compliance Portal: Salesforce-owned Model Cards · 1 September 2026 · primary source
- Salesforce AI Research / Hugging Face: Model Card Salesforce/xLAM-7b-r · 3 November 2026 · primary source
- Joseph Saveri Law Firm: Salesforce Large Language Model Litigation – Case Background · 15 October 2025 · secondary source
4 · Why the harm rule does not apply
The harm rule does not apply: in the period there is no judgment, no authority decision and no penalty concerning harm from Salesforce's AI. The prompt injection flaws were fixed, and the lawsuits have produced no findings.
5 · Controversies (7)
- mediumWith SalesBleed, Zenity Labs showed zero-click data leakage from Agentforce via web forms; Salesforce fixed the flaws by August 2026. · 24 September 2026 · abgeschlossen
- mediumAccording to the lawsuit, Salesforce used copyrighted books to train its xGen models; the cases were consolidated in 2025 and are pending. · 15 October 2025 · open
- mediumWith ForcedLeak, Noma Security showed that Agentforce could leak CRM data via prompt injection; Salesforce fixed the flaw. · 25 September 2025 · abgeschlossen
- mediumA vishing and token campaign stole data from hundreds of customer instances; class actions are pending, and the JPML sees no platform flaw. · 4 June 2025 · open
- minorMore than 1,400 employees called for alleged AI offers to ICE to be dropped; AI agents also received approval at the US Department of Defense. · 10 February 2026 · unklar
- minorPromptArmor showed that Slack AI could disclose data from private channels via prompt injection; Slack subsequently released a patch. · 20 August 2024 · abgeschlossen
Further controversies (1)
6 · What remains open
- Independent audit of the AI disclosures in the sustainability and Trusted AI reports
- Metrics on red-teaming, AI incidents and enforcement of the AI acceptable use policy
- Complaint channels for external people affected by AI outputs
- Approval and monitoring of the third-party models used
- Certification body and exact scope of the ISO 42001 certification
- Full text of the second Trusted AI & Agents Impact Report 2026
7 · How it was assessed
Cut-off date: 3 November 2026 · 41 searches, 38 fetches.
Primary documents read (11)
- Form 10-K FY 2026 (main text with read-more, XBRL R26 Cybersecurity, R23 Legal Proceedings)
- Form 10-Q as at 2026-07-31 (selected sections, XBRL R19 Legal Proceedings)
- Proxy Statement DEF 14A 2026 (first 100,000 characters)
- Form 8-K Annual General Meeting 2025
- Trusted AI and Agents Impact Report, February 2025
- FY26 Stakeholder Impact Report (first 100,000 characters)
- ISO 42001 entry in the Compliance Portal
- Model cards overview in the Compliance Portal and model card xLAM-7b-r
- Agentforce Developer Guide: Trust Layer
- JPML order MDL 3164/3170 dated 2025-12-16
- Joint Case Management Statement, In re Salesforce LLM Copyright Infringement Litigation (ECF 45)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0