When the Simple Attack Wins
Sometime before Monday, attackers figured out something the AI security industry had largely talked around: you don’t need a superintelligent hacking model to cause real damage. Exploiting Meta’s AI customer support agent, they asked it to link Instagram accounts to email addresses they controlled. The agent complied. Accounts were stolen. The method required no advanced exploit, no sophisticated prompt injection chain, no model with autonomous offensive capabilities – just a direct request to a system that wasn’t built to say no.
The incident reframes a conversation that had, until now, orbited almost entirely around Anthropic’s Mythos model.
Anthropic had withheld Mythos from general release after determining it was too capable at hacking for safe public deployment. That decision, reasonable on its face, trained public attention on the threat of AI systems that could overwhelm infrastructure at scale. The Instagram exploit is a counter-argument in the form of a real event: as companies push more customer-facing work onto AI agents, the surface area for low-sophistication attacks grows alongside the tools themselves.
Anthropic Pushes for a Global Pause – Skeptics Push Back
Separately, Anthropic called this week for a coordinated global slowdown in AI development, flagging the risk of models reaching the point of self-improvement. The company wants an international plan to stop that trajectory before it becomes unmanageable. Reuters reported on the coordinated response Anthropic is lobbying for, while the Wall Street Journal covered its self-improvement concerns in detail.
The timing drew immediate skepticism. The Register noted that a leading AI lab calling for a global slowdown – while continuing to build and ship its own models – carries a certain strategic convenience. Whether the call is principled, self-interested, or both, it lands in a week already crowded with AI governance signals: the White House held talks about the government acquiring financial stakes in AI firms, a move Sam Altman reportedly pitched to the administration last year. South Korea’s labour minister, Kim Young, proposed that tech firms share AI profits with staff and suppliers – a position that gained weight after he helped avert a major strike over AI profit-sharing at Samsung.
Canada launched its long-anticipated AI strategy, promising over $2 billion in funding and a target of 250,000 new jobs. The policy ambition is notable; so is the gap between funding announcements and the actual delivery of jobs in sectors where AI is simultaneously automating work. The relationship between AI investment and net employment has rarely been straightforward.
What Constant AI Use May Be Doing to Human Thinking
Gloria Mark, a psychologist at the University of California, Irvine, has spent years tracking how digital tools affect cognition. Her research shows attention spans have fallen sharply over time – a decline she links to higher stress and measurably lower performance. Now she’s extending that concern to AI assistants specifically. “You’re deferring your cognitive work to AI,” Mark said. “And it’s not good for us.” Her argument is that tools like ChatGPT and Claude may weaken critical thinking and emotional intelligence not through any dramatic intervention, but through gradual disuse – the same way any capacity atrophies when left idle.
Mark believes course-correction is possible, contingent on people changing their relationship with these tools rather than simply using them less. The distinction matters: the question isn’t abstinence from AI, but whether users stay cognitively active during interactions or outsource thinking entirely. That’s a harder behavioral shift than it sounds, especially as AI interfaces are designed to minimize friction and maximize reliance.
The cognitive concern sits alongside a separate but related development: the White House is pushing to bring AI into American healthcare, with chatbots diagnosing illness and prescribing medicine. Whether that increases or decreases the cognitive load on patients and doctors is an open question – one that exists independently of whether healthcare AI actually improves patient outcomes, which, as MIT Technology Review has reported, remains unresolved.
Bot Traffic Now Outpaces Human Traffic Online
Cloudflare reported this week that 57.4% of web traffic now originates from bots, meaning automated systems have crossed the threshold to become the dominant presence on the internet. The milestone arrived ahead of schedule – Cloudflare’s own CEO had predicted it would happen by the end of 2027. The early arrival reflects how quickly AI-driven automation has scaled across data collection, indexing, scraping, and content generation pipelines.
Elsewhere in the week’s technology news: scientists used a newer gene-editing technique to precisely edit human embryo genes for the first time, a development the Guardian called a step toward genetically modified babies. Meta quietly embedded facial recognition code into its app for smart glasses – an exploratory feature that would identify people through biometric data. Investment in agricultural technology is growing at a time when food market volatility is putting pressure on supply chains globally. And new research found that bumblebees can use tools to solve problems, which has nothing to do with AI but is, by any measure, a better piece of news than most of what surrounds it.
The Cloudflare CEO who set a 2027 target for bot traffic dominance watched it happen in mid-2025 and responded: “Welp, that happened faster than I predicted.” It’s a line that applies, with varying degrees of alarm depending on who’s reading it, to almost everything else that happened this week.
