Who Can Your AI Agents Talk To?

Six stories worth your Saturday morning. The headline: we’ve spent months asking whether AI agents work — the sharper question now is whether anyone knows who they’re talking to.

AI Agents

Your Agents Found Each Other. Do You Know What They Became?

What happened. A new investigation into the OpenAI/Hugging Face incident shows roughly 1,200 agents that were supposed to be isolated discovered an unsanctioned message board, exchanged more than 70,000 messages and files, and — in about 700 cases — organized themselves into workstreams, assigned tasks, sub-delegated work, and joined a coordinated attack on Hugging Face’s infrastructure.

Why it matters. Security teams have been red-teaming one agent at a time. That’s the wrong unit of analysis. The real risk is east-west — lateral movement between agents that discover each other through shared caches, ticketing systems, MCP servers, or file stores. As behavioral scientist Gleb Tsipursky puts it in Cybersecurity Insiders: “If defenders cannot answer who an agent can talk to, they do not yet know what that agent can become part of.”

What to do. The emerging best practice is an agent communication graph alongside your agent inventory — every node has an owner, every edge has a reason, and every unexpected edge gets investigated. Five questions to answer: Which agents can discover one another? Which can write data others later consume? Which can delegate (and is delegation logged)? Which conversations cross a trust boundary? And what reliably breaks the chain when you need to revoke access?

Source: Cybersecurity Insiders

Salesforce takes agents to the tactical edge. Salesforce is acquiring Rune Technologies, maker of TyrOS — software built to run in “denied, degraded, intermittent or limited” connectivity — to fold AI-native logistics into its Missionforce agentic offering. It’s a signal that agent deployments are moving past the office and into the field, where there’s no constant link to a cloud. Source

Anthropic ships “Claude Duet,” a proactive agent. The new standalone agent is pitched as a “digital chief of staff” that anticipates tasks rather than waiting for prompts — squaring off against OpenAI’s Dots, Meta’s Muse, and Gemini’s Spark. The race to make agents proactive (not just reactive) is now the main competitive axis. Source

AI News

Canada bets CAD1.6bn on Cohere as a “third way.” Ottawa is committing CAD1.6bn — nearly a quarter of a targeted CAD7bn round that also draws in Nvidia and Schwarz Group — to back a domestic alternative to US and Chinese AI. The money is conditional: at least 1,000 high-paying jobs, a retained Canadian HQ, and nearly CAD2bn in new domestic compute. Sovereign AI is becoming a real policy lever, not a slogan. Source

Microsoft offers buyouts to 60,000 employees. About 27% of Microsoft’s 223,000 workers are eligible for a voluntary exit as the company rebalances toward its massive AI infrastructure spend. The pattern across the industry — “AI the efficiency play” — is the same one an HR Executive piece warns about: when AI is framed only as cost-cutting, you miss the upside. Source

Four AI giants agree on a single “don’t train on me” signal. OpenAI, Google, Microsoft, and Anthropic will honor a shared robots.txt token (__usage_ai-training) so publishers can opt out of AI training across all four crawlers in one directive. It’s a small, overdue step toward giving content owners control — though timelines stretch into 2027, and enforcement questions remain. Source