Toronto-based AI company founded 2019 by ex-Google researchers. Builds secure, enterprise-grade LLMs for regulated industries like finance/healthcare. Focus: customizable, on-prem deployable models enhancing productivity via agentic AI like North platform.Cohere About, Wikipedia
Toronto-based AI company founded 2019 by ex-Google researchers. Builds secure, enterprise-grade LLMs for regulated industries like finance/healthcare. Focus: customizable, on-prem deployable models enhancing productivity via agentic AI like North platform.Cohere About, Wikipedia
Command R+ supports multi-step tool use/agents, RAG w/ citations/grounding, improved tool decision-making (Aug 2024 update). Enables reasoning/dynamic actions over tools for production workflows.Cohere Docs, Models
Feb 2025: Sued by Condé Nast/McClatchy for copyright infringement (reproduced full articles, hallucinated fake news).TVNewsCheck Aug 2024: Jailbreak prompt bypasses filters.GitHub No data breaches/leaks found.
Detailed breakdown of every risk category for enterprises deploying Cohere models in agentic AI workflows.
Can generate fabricated content attributed to sources (e.g., fake Guardian article in lawsuit). RAG mitigates but not eliminates; acknowledged business risk.Lawsuit, Cohere Blog
SOC2 Type II, GDPR DPA w/ EU-US TIA, opt-out training, VPC/on-prem deploys, encryption. No incidents; general API key/prompt injection risks (mitigated e.g., 86% on Aya).Security, Trust Center
Training data biases (race/gender/etc.); BOLD dataset eval. Potential discriminatory outputs in enterprise use.Model Card, SaferAI
Jailbreaks via persona prompts (2024); prompt injection reducible. Safety modes/post-deploy backdoors noted.GitHub, Hirundo
Agentic tool chains vulnerable post-jailbreak, potentially executing unintended API calls despite decline training.Agent Jailbreak
Regular updates (e.g., R+ 08-2024: +50% throughput, better instructions). Leaderboard selection bias noted in Cohere paper.Models
GDPR/CCPA compliant; ongoing copyright suit. EU AI Act high-risk for agents.Trust Center
2025 lawsuit: trained on 4K+ unauthorized works, outputs full paywalled articles.TVNewsCheck
Text-focused; low native deepfake risk. Jailbreak-enabled harmful text gen possible.SaferAI
E&O for hallucinations/agent errors; cyber for breaches/injection; D&O for fines/litigation. Emerging AI policies; check exclusions.eSpecialty, Resultsense
RBC, LG, McKinsey, Salesforce, Nvidia, Oracle, Notion, Google.CBInsights, Video