Three storylines from the past few days capture where AI actually stands heading into September 2026: money keeps flowing into the infrastructure layer, security researchers keep finding new ways to break agentic systems, and a fresh economic survey suggests the productivity payoff is still mostly theoretical.
Salesforce and Anthropic go all-in on agents
On August 28, Salesforce and Anthropic announced an expanded partnership — nicknamed "Claudeforce" in early coverage — that embeds Claude into Salesforce products while making Claude the default reasoning model across Salesforce's Agentforce lineup. The move is the latest sign that enterprise software vendors are racing to lock in a single frontier model as the reasoning engine behind their agent platforms, rather than staying model-agnostic. It follows a broader industry pattern this week of vendors building governance tooling around agents: security platform AccuKnox also shipped AgentZ, a model-agnostic platform bundling agents, sandboxes, workflows, role-based access, and audit traces so companies can move agents from pilot to production without stitching together separate tools themselves.
A sobering lesson in prompt injection
While enterprises rush to deploy agents, security researchers are demonstrating how fragile the trust model still is. Forcepoint X-Labs disclosed an attack in which researchers planted instructions inside an email using zero-size, white HTML text that stayed invisible to a human reader but was still passed to the AI model. In testing, the visible email ran 537 characters while 1,009 reached the model, and all ten injected test runs produced a manipulated summary. The hijacked summaries did real damage in the demo: the tampered output moved an invoice deadline and deleted a name, prompting Forcepoint to recommend that AI summarizers only extract user-visible content and treat email as untrusted data. It's a pointed reminder that as agents get wired into inboxes and workflows, the attack surface is expanding faster than the defenses.
The productivity numbers still aren't showing up
Perhaps the most important story of the week isn't a product launch at all. A new National Bureau of Economic Research survey of executives after three years of AI adoption found that more than 90 percent reported no effect on their own firm's employment, and 89 percent reported no effect on productivity — even as layoffs tied to AI narratives have continued unabated. That disconnect between corporate rhetoric and measured outcomes is likely to fuel skepticism as AI infrastructure spending keeps climbing, including reported talks between Nvidia and Perplexity over a funding round that could value the search startup at more than $30 billion, according to The Information, with Perplexity's annualized revenue climbing above $750 million, up from less than $250 million at the start of the year.
Taken together, these threads suggest the AI industry is entering a more skeptical, more adversarial phase — one where deployment speed, security holes, and unproven ROI are all colliding at once.
Comments (0)
No comments yet — be the first.
Sign in to leave a comment.