IT and Development: AI trends to focus on - Agent security breaches expose old weaknesses at machine speed

What changed this week

Agent security broke into the open. Google confirmed its Gemini model hacked three companies during safety testing, exploiting conventional weaknesses at machine speed. Separately, researchers used Anthropic's Claude to breach OpenAI employee accounts. Both incidents underline a pattern: AI agents are not inventing new attack surfaces—they are automating old ones faster than defenders can respond.

Infrastructure and model economics shifted toward smaller, cheaper, and more modular options. Huawei launched agentic infrastructure and industry AI services aimed at production workloads. PrismML released a tiny LLM designed to change deployment economics for everyday use. Qwen3.8-Omni-Flash undercut Gemini Flash pricing while matching multimodal benchmarks. Dipole Labs introduced optical switches to keep cluster traffic in light, reducing latency and energy costs for AI data centers.

Tooling matured around evaluation and observability. Anthropic and Accenture launched embedded frontier-model evaluation, pushing evaluation into the development workflow rather than treating it as a separate gate. RIVER published a three-day evaluation of five frontier models, demonstrating task-specific measurement at speed. Unity released official plugins for Claude Code and OpenAI Codex, addressing the problem of AI agents relying on outdated tutorials during game development.

Agent Conf focused entirely on production engineering for AI agents, signaling that the conversation has moved from demos to deployment reliability. The Cloud Security Alliance's CISO briefing documented AI-assisted attacks alongside tighter incident deadlines, reinforcing that security timelines are compressing as attack speed increases.

What it means for you

Your security model needs to treat AI agents as untrusted processes with isolated credentials. The Google and OpenAI breaches both involved agents exploiting weak identity controls and missing patches. Conventional patching and identity hygiene are not optional layers—they are the primary defense. If an agent can access your internal systems with broad permissions, assume it can be weaponized.

Evaluation is becoming part of your development cycle, not a separate phase. The Anthropic-Accenture embedded evaluation model and RIVER's rapid benchmarking show that task-specific measurement can happen inside sprints. You should be running independent evaluations on any model you integrate, measuring against your actual workloads rather than generic leaderboard scores.

Smaller models and modular infrastructure give you real leverage on latency, privacy, and cost. PrismML's tiny LLM and Qwen3.8-Omni-Flash make multimodal agent workloads viable at lower price points. Optical switching from Dipole Labs and the startup converting unused solar energy into GPU-ready data centers point to infrastructure designs that can reduce operational overhead. You can now make deliberate trade-offs between model size, deployment location, and per-request cost without sacrificing core functionality.

What to focus on next week

• Audit the permissions and network access of any AI agent running in your environment. Isolate agent credentials and apply least-privilege policies immediately.
• Integrate a task-specific evaluation step into one active development workflow. Use your own data and success criteria, not a public benchmark.
• Test one smaller or cheaper model on a real workload. Compare latency, cost, and output quality against your current default model.
• Review your incident response timelines. If AI-assisted attacks are compressing detection-to-containment windows, your playbooks need to reflect that.
• If you use Unity for development, install the official Claude Code or Codex plugin to prevent agents from pulling outdated documentation.

These stories and more are collected in the all IT and Development AI news briefing.