GPT-Red, bioresilience and the real problem with agents

Today’s most important AI news

Thursday arrived loaded with news revealing a central tension in the industry: the distance between what AI promises and what it can actually do today. OpenAI presented an unprecedented strategy to harden its models against cyberattacks, Google DeepMind formalized its stance on one of the most serious risks of using AI in biology, and a study of 101 companies confirmed what many suspect but few say out loud about enterprise agents. It wasn’t a day of spectacular launches, but of honesty about current limits — and that, for anyone making product or technology decisions, is worth more.

GPT-Red: OpenAI built an AI hacker to protect its own models

OpenAI trained a specialized model called GPT-Red whose sole purpose is to attack the company’s other models. The logic is straightforward: instead of waiting for outside actors to find vulnerabilities, GPT-Red looks for them first in a systematic, automated way. The model was used as a “sparring partner” during the training of GPT-5.6, launched last week, which OpenAI describes as its most robust release to date against cyberattacks.

What makes this approach notable is that it isn’t traditional human red teaming — that process where security specialists try to trick the model manually — but automating it at scale. GPT-Red can generate and test attack vectors continuously, something a human team can’t match in volume or speed. It is, essentially, applying the logic of software fuzzing to language models.

For teams working with models in production, this points to a clear direction: LLM security evaluation will require AI tools, not just human auditors. If your organization is deploying its own or fine-tuned models, the question is no longer whether you’ll do red teaming, but when you’ll start automating it.

Google DeepMind and the risk nobody wants to ignore: bioresilience

Google DeepMind and Isomorphic Labs published their joint approach to bioresilience, a term describing the ability to anticipate and mitigate the risk that AI models — especially those designed for biology and drug discovery — are used to create or enhance pathogenic agents. It’s not a theoretical topic: AlphaFold and its successors can already predict protein structures with a precision that was unthinkable a few years ago, which opens doors both for medicine and for deliberate harm.

The document doesn’t detail all the technical mechanisms, but the fact that DeepMind publishes it as formal policy is significant. The company is explicitly acknowledging that its AI tools for the life sciences have a dual-use risk profile — that is, the same knowledge that accelerates drug design can, in the wrong hands, accelerate the design of something much more dangerous.

For developers and PMs working in biotech or life sciences with AI, this marks the start of a regulatory conversation that will grow. Access restrictions, usage controls and purpose documentation will become standard requirements, not optional ones. Anyone building without thinking about this today will have to redesign later.

The real problem with agents: most are chatbots with better marketing

A VentureBeat analysis of 101 companies deploying “AI agents” found something many in the industry already sensed: the vast majority of what’s sold and implemented as an agent is, in reality, a chatbot with a more elaborate workflow. Real agentic orchestration — systems that execute multiple steps autonomously, handle errors, coordinate tools and make decisions in real time — remains the exception, not the rule.

The study also reveals that platform consolidation is happening faster than expected: Anthropic’s Claude leads among the companies surveyed, chosen mainly for its reliability in executing multi-step tasks. However, the same companies are deliberately designing their control layers in a hybrid way to avoid lock-in with a single vendor. And token spend control — something critical for anyone running agents in production — remains a capability most haven’t implemented.

The practical implication is direct: before investing in complex agentic architectures, it’s worth honestly auditing what’s actually running in production. If the system can’t execute a five-step task without human intervention, it isn’t an agent — and treating it as one creates expectations the system can’t meet. The problem these companies identify isn’t about the platform or the model; it’s about deployment, clear definition of use cases and realistic success metrics.

In a nutshell

Today’s three stories point in the same direction: the industry is entering a phase of forced maturation. OpenAI automates its own red teaming because the volume of attacks demands it. DeepMind publishes biosecurity policies because regulators — and its own board — are pressing it to. And companies are discovering that calling a chatbot an “agent” doesn’t make it one. There’s a pattern here: the hype of the last two years is colliding with the operational and security limits of the real world, and the organizations that will come out best are those being honest about that gap before someone else forces them to be.

---

Sources used: MIT Technology Review — GPT-Red (https://www.technologyreview.com/2026/07/15/1140514/meet-gpt-red-an-llm-super-hacker-openai-built-to-make-its-models-safer/), Google DeepMind — Bioresilience (https://deepmind.google/blog/our-approach-to-bioresilience/), VentureBeat — Agentic Orchestration (https://venturebeat.com/ai/agentic-orchestration-enterprise-ai-organizations-have-a-deployment-problem-not-a-platform-problem-and-most-are-calling-chatbots-agents)