AI agents make real decisions: markets, drugs and security

Today’s most important AI news

Thursday was marked by a trend that’s increasingly hard to ignore: AI agents are no longer lab prototypes, but systems that make decisions with real economic consequences. Today’s news ranges from agents negotiating freight in transport markets, to security vulnerabilities in multi-agent systems and the use of AI to design the next generation of biologic drugs. The pattern is clear: the conversation has stopped being “can AI do this?” and become “what happens when it’s already doing it?”

When GPT, Claude and Gemini negotiate freight: what happens to markets

A paper published today on arXiv presents something that should interest anyone operating in logistics or marketplace platforms: researchers simulated a freight transport market where fifty buyer agents — built on OpenAI’s GPT, Anthropic’s Claude and Google’s Gemini — negotiated truck capacity over thirty simulated days. The results are uncomfortable.

What they found is that delegating carrier selection to LLM agents significantly concentrates the market. The agents tend to favor the same carriers repeatedly, partly because the models share training biases and partly because the design of the information they receive — which options the platform presents to them — almost completely determines their choice. In other words, whoever controls the data feed controls the market, even if no human is making the final decision.

The practical implication for product managers and platform founders is direct: if you’re building a marketplace where buyers will eventually delegate decisions to agents, the design of how information is presented isn’t a UX detail, it’s the variable that determines who wins. Existing antitrust regulations aren’t designed for this scenario, and the paper concludes that new platform design rules are needed before this concentration becomes structural.

Multi-agent systems have a security problem nobody is measuring well

A paper published today under the name ChannelGuard documents something security teams should read this week: in systems where multiple LLM agents are chained together — a planner, executor agents, a verifier, a synthesizer — every connection between agents is an unmonitored channel through which an attacker can inject malicious instructions. Current defenses only protect the system’s initial entry point, not what happens between agents.

The researchers tested this against 2,100 scenarios and found that prompt injection attacks that fail at the system’s external boundary can succeed when aimed at the internal channels between agents. It’s an attack vector that Llama Guard filters, perplexity-filtering systems and SmoothLLM solutions simply don’t cover because they were designed for a different threat model.

For any team putting agents into production today — especially in flows that touch sensitive data or execute actions with real effects — this is a concrete warning sign. The fact that each individual agent passes safety evaluations doesn’t guarantee the composite system is safe. The chain of trust between agents needs its own layer of defense, and today that layer doesn’t exist as standard on any of the major platforms.

AI is redesigning how biologic drugs are created

MIT Technology Review published an analysis today of how AI is transforming the design of biologic drugs — therapies based on engineered proteins, not traditional synthetic chemistry. The conventional process takes years and most candidates never reach patients. AI is changing that equation at two specific points: designing the protein structure and predicting which candidates are most likely to survive the clinical phases.

What makes this moment different isn’t that AI “helps” scientists search existing libraries of compounds, but that it’s generating new protein structures no human team would have explored, and then predicting their behavior before synthesizing them physically. That compresses the experimentation cycle from months to days in some stages of the process.

For executives in Latin American healthtech or biotech, the relevant question isn’t whether this is happening — it’s happening in the labs of Genentech, Recursion and several European startups — but when the cost of accessing these tools drops enough for smaller teams to compete. Protein models like AlphaFold are already accessible; the next step is generative design tools, and that cycle is shortening fast.

In a nutshell

What connects today’s three stories is the same tension: the speed at which AI agents are taking on real responsibilities exceeds the speed at which frameworks to contain them are being built. In logistics, agents are already concentrating markets without clear rules about how options should be presented to them. In security, multi-agent systems are being deployed with an exposed flank nobody is systematically measuring. In medicine, AI designs molecules that could become treatments, and the regulatory conversation is just beginning. The “deploy first, regulate later” narrative that defined the era of consumer platforms seems to be repeating itself with AI agents, but with potentially more immediate and less reversible consequences.

---

Sources used: https://arxiv.org/abs/2607.19967, https://arxiv.org/abs/2607.19430, https://www.technologyreview.com/2026/07/23/1140346/how-ai-helps-scientists-design-the-next-generation-of-medicines/