Here’s our latest Yes-You-Should-Worry-About-AI roundup. Salesforce research shows AI agents exhibit “near-zero confidentiality awareness” when using customer data, while Anthropic research shows AI models can often sabotage user intentions without being caught by the AI models that are supposed to monitor them. And Asana found the much-touted Model Context Protocol exposed Asana data to other MCP users.
Continuing to enforce only laws he likes and ignore the rest, President Trump has further extended TikTok’s ability to operate despite a U.S. law requiring that it be sold or shut down six months ago. You may recall that the pre-election panic over TikTok sharing U.S. data with the Chinese government somehow vanished after the President learned GOP mega-donor Jeff Yass held a major TikTok investment.
Is AI working for humans or are we working for AI? Let’s just note that the goal of Adobe’s new LLM Optimizer (no release date available) is helping a brand’s content appeal to LLMs like ChatGPT and Perplexity. More evidence that agentic traffic is gaining importance compared with traditional search.
The web search tool Perplexity looks like it’s taking on Google with its latest release, a web browser named Comet. Comet claims to move “from browsing to thinking,” surfacing information and search results in response to natural language queries. Comet is available immediately to Perplexity Max subscribers; others can join a wait list.