Millions of AI agents: the new problem nobody has solved

Today’s most important AI news

Thursday was dominated by two warning signs pointing in the same direction: the industry is building increasingly complex AI systems without having solved the coordination and reliability problems that emerge at scale. Google DeepMind raised its hand about the risks of millions of agents interacting without human supervision, while researchers showed with concrete benchmarks that frontier LLMs still aren’t ready to operate on their own in cybersecurity. These aren’t academic topics: they’re limits already affecting product and architecture decisions.

Google DeepMind warns about the chaos of agents en masse

Google DeepMind is funding research on a scenario that until recently seemed like science fiction but is becoming operational: what happens when millions of different AI agents — built by different companies, with different goals — start interacting with each other on the internet without direct human supervision. Rohin Shah, director of AGI safety and alignment research at DeepMind, described the problem as one of the hardest emerging risks to anticipate precisely because it doesn’t come from a single system but from the interaction between many.

The critical point here isn’t a rogue agent acting alone, but the collective behavior that emerges when systems designed to follow other agents’ instructions are chained together in complex pipelines. An agent can behave perfectly well within its domain and still take part in cascades of decisions nobody designed and nobody controls. It’s the hardest coordination problem in applied AI: there’s no single responsible actor, and the effects are diffuse.

For developers and PMs building multi-agent architectures today, this has immediate implications. Before scaling a system where agents call each other or execute tasks on behalf of other agents, it’s worth asking: who audits the intermediate decisions? Is there an interrupt mechanism when the pipeline goes off course? DeepMind hasn’t published solutions yet, but the mere fact that it’s funding this line of research signals that the problem is real and urgent.

Frontier LLMs still aren’t reliable for real cybersecurity

A dual benchmark published on arXiv evaluated whether the most advanced language models available — including GPT-5.4, Claude Opus 4.6, Sonnet 4.6 and Gemini 3.1 Pro — are ready to operate in cybersecurity. The results are sobering: none reached the level needed to replace specialized processes. The benchmark combined source-code-level vulnerability detection in C, Java and Python with security testing on real web applications with 118 documented vulnerabilities spread across more than 20 CWE families.

A second study, also published this week, evaluated whether agents based on open-source models running in Ollama could replace static application security testing (SAST) tools. The conclusion was similar: generative agents still produce too many false positives and have insufficient recall to be used as the sole detection mechanism in production environments. They have value as an additional layer, not as a substitute.

What both studies suggest isn’t that AI is useless in security — there are already clear use cases in triage and assistance — but that the gap between “impressive in a demo” and “reliable in production” remains enormous in domains where errors have real consequences. If you’re evaluating adding LLMs to your security pipeline, the message is clear: use them to accelerate human analysts’ work, not to eliminate that layer.

OpenAI and Oracle: enterprise infrastructure consolidates

OpenAI announced that its models, including Codex, are now accessible directly through existing Oracle Cloud spending commitments. In practical terms, this means companies with active Oracle contracts can consume OpenAI’s API and deploy models without opening a separate commercial relationship, and with the governance and security controls Oracle already offers its corporate customers.

The move is part of a broader pattern: the big AI providers are looking to reduce enterprise adoption friction by integrating with the purchasing cycles large organizations already have in place. For a Latin American company with an Oracle presence, this significantly lowers the barrier to entry, both in procurement and in compliance.

The relevant detail for technical teams is that this doesn’t change the models’ capabilities, but it does change the conversation with procurement and legal. Having AI consumption within an existing enterprise contract simplifies internal approval in organizations where adoption speed has been limited more by processes than by technical will.

In a nutshell

Today’s pattern is that of an industry that accelerated deployment before solving the problems of scale. DeepMind acknowledges it doesn’t fully understand what happens when its own agents interact en masse; cybersecurity benchmarks show that the world’s best models fail under real conditions; and meanwhile, OpenAI and Oracle work to make it easier for companies to adopt those same models at enterprise scale. It’s not a contradiction — it’s the normal speed of technology — but it is a signal for decision-makers: today’s risk doesn’t come from AI that doesn’t work, but from AI that works well 90% of the time and nobody knows what it does in the remaining 10% when systems are chained together.

---

Sources used: https://www.technologyreview.com/2026/06/11/1138794/google-deepmind-is-worried-about-what-happens-when-millions-of-agents-start-to-interact/, https://arxiv.org/abs/2605.23243, https://arxiv.org/abs/2606.11672, https://openai.com/index/openai-on-oracle-cloud