Grok 4.7 lands at $2 per million tokens as AI agents outrun safeguards

The most unsettling story today isn't a model launch. It's a Hong Kong IP address with an open directory. CrowdStrike found what looks like the first concrete evidence of an autonomous AI pentesting agent being turned loose on South Korean banks, and it found it only because the operator forgot to close their sessions. I think that detail matters more than the attack itself. We have been discussing agentic threats as a theoretical category for two years, and the first hard evidence arrived through sloppy operational security, not through any defensive system catching it. How many operators weren't sloppy?

That sits uncomfortably next to the broader argument that agents are gaining autonomy faster than safeguards can follow. I find the framing slightly too tidy, because it implies a race between two comparable things. Capability is cheap to scale and easy to demonstrate. Controls are slow, unglamorous and hard to sell. Microsoft at least seems to understand this: its latest announcements focus on protecting Windows devices from the risks agents themselves pose, alongside the new Copilot+ PCs. Notice the order of priorities. The company that wants agents on every desktop is now spending real effort on defending the desktop from them.

Meanwhile the capability side keeps moving. Amazon's AGI Lab is betting that the next leap for agents isn't better clicking but actually perceiving screens the way humans do. That is a genuinely interesting shift, though it cuts both ways. An agent that understands interfaces rather than scripting against them is far more useful for legitimate automation, and it is also far harder to detect and block when someone points it at a bank. By the way, this is exactly the kind of second-order effect that rarely makes the product announcement.

On the commercial front, xAI shipped Grok 4.7 at $2 per million tokens, nearly doubling its Terminal-Bench score and benchmarking itself against Anthropic's Claude Sonnet 5.5 and Mistral Large 4. The price is the real story. Benchmarks are self-selected and I would wait for independent testing before drawing conclusions, but aggressive pricing puts pressure on everyone, and cheap capable models mean more agents, run by more people, with fewer questions asked. Microsoft's Jacob Andreou made a related point about rebuilding Copilot from scratch: the gap between what models can do and what products deliver is widening. I would add that the gap between what models can do and what we can safely govern is widening too.

That is the thread I would pull on this week. Product teams are racing to close the usability gap, while the safety gap gets discussed in interviews, including DeepMind COO Lila Ibrahim's reflections on existential risk, rather than closed in code. If you are building with agents, the practical lesson from Hong Kong is simple: assume someone else is running the same technology with fewer scruples, and design your logging, permissions and session handling accordingly. The next CrowdStrike discovery may not depend on a careless attacker.