join.dev
AI news · 10 stories
Cohere launches North 2 with agent memory and cost controls · GitHub launches ReviewBench for evaluating AI code review · TikTok introduces an AI shopping assistant and direct checkout
Cohere introduced North 2 with cross-session memory, reusable skills, redesigned orchestration and granular spending controls. Private cloud, on-premises and air-gapped deployment options give enterprises more control over where their agents operate.
Read at Cohere
GitHub released ReviewBench, an open benchmark using representative pull requests, combined reference findings and a consistent evaluation rubric. It aims to compare what AI reviewers catch and miss while measuring the tradeoff between useful findings and noise.
Read at GitHub
TikTok announced Shopping Assistant for conversational product guidance and Buy Direct for purchases without leaving the app. The offerings connect product discovery, shopping questions and checkout within TikTok, with commerce and payment partners supporting the rollout.
Read at TikTok
Google Research published a workshop report on unresolved privacy and security problems for autonomous agents. It calls for defenses spanning systems, models, users and the wider ecosystem, with agent behavior constrained by the context in which it acts.
Read at Google
The IWF assessed 6,310 AI-generated child sexual abuse images in the first half of 2026, exceeding its full-year 2025 count by 40%. It urged EU policymakers to give platforms a lasting legal framework for detecting and removing this material.
Read at Internet Watch Foundation
A new Anthropic case study explains how Cresta uses the Claude Agent SDK beneath Conductor to build, evaluate and improve customer-experience agents. Cresta reports that early deployments took roughly half as long, while emphasizing policy controls and ongoing testing.
Read at Anthropic
Google documented a beta tool that creates localized copies of Search campaigns using generative AI for assets, targeting and optional landing-page translation. Advertisers can review translated assets before publishing, reducing manual localization work while preserving the original campaign.
Read at Google
Cisco published a comparison of Jev and Laya decision models with a classifier trained for its content-safety taxonomy. Its results favored specialized training at strict false-positive thresholds, while calling for stronger held-out testing before drawing deployment conclusions.
Read at Cisco
Cohere and PwC announced a global alliance launching first in Canada, combining Cohere models and North with PwC implementation and governance expertise. The partnership targets moving enterprise AI pilots into production while retaining control of sensitive data and infrastructure.
Read at Cohere
Cisco described how its internal RunOn platform connects cloud spending, ownership and operational findings with AI queries and supported cloud actions. Integration with Cisco’s MyAgent remains planned, illustrating how organizational context can guide cloud operations.
Read at Cisco