Executive Intelligence Brief - Week 13, 2026

A startup founder told us he replaced his three junior developers with an AI coding agent. The agent costs $500 a month total, versus $30,000 a month for the team. The code quality is comparable. We explored both the exciting and deeply uncomfortable implications of this shift.

New research from ETH Zurich demonstrates prompt injection attacks that bypass all known defensive measures with 97% success rate. As enterprises connect LLMs to internal tools and databases, the attack surface expands dramatically. We need to treat LLM-connected systems with the same security rigour as we treat database-connected web applications.

We are starting to see a new category of tech debt: AI-generated code that nobody on the team fully understands. It works, it passes tests, but when it breaks nobody knows why. Engineering leaders need to think about this before rolling out autonomous coding tools org-wide.

We are advising all our portfolio companies to establish AI red teams. The threat surface of LLM-powered applications is fundamentally different from traditional software. Prompt injection, data poisoning, model theft, and adversarial inputs require specialised security expertise that most organisations lack.

After two years of skepticism about AI application companies, we are seeing clear evidence that the application layer is working. The best AI applications combine three things: a specific workflow (not a horizontal tool), proprietary data flywheels, and AI-native UX that makes the old way feel broken.

GPT-4o, Claude 3.5, and Gemini 1.5 Pro have all reached the point where their vision capabilities are production-ready for enterprise use cases. Document understanding, visual QA, and image-to-structured-data extraction now work reliably enough for automation. I expect multi-modal to become the default modality for enterprise AI within a year.

Hot take: the companies failing with AI agents are the ones treating them like software. Agents need operating procedures, escalation paths, and supervision models -- just like human employees. The orgs succeeding are the ones applying operational management thinking, not just engineering thinking.

The number one mistake we see in production RAG systems is poor chunking strategy. People use fixed-size chunks because it is easy, but semantic chunking -- where you split on topic boundaries -- improves retrieval accuracy by 15-25% in our benchmarks. The second mistake is not having an evaluation framework before you start.

This week's biggest development: Cognition's Devin 2.0 achieved 58% on SWE-bench Verified, up from 43% six months ago. Meanwhile, OpenAI's internal coding agent reportedly resolves 70% of internal bug reports autonomously. The era of autonomous code generation is arriving faster than most engineering leaders expected.