AI & Automation Briefing - July 10, 2026
Prompt Injection Flaw in GitHub Agentic Workflows Exposed Private Repository Data
Noma Labs researchers discovered a critical prompt injection vulnerability in GitHub Agentic Workflows, the platform's new system that pairs GitHub Actions with an AI agent powered by Claude or GitHub Copilot. The attack, which Noma named GitLost, required no authentication. An attacker could post a crafted issue in any public repository belonging to an organization, and the GitHub agent would read the issue body, treat the embedded instructions as legitimate, and exfiltrate data from private repositories in the same organization. The root cause is a missing trust boundary between system-level directives and untrusted user-generated content. The vulnerable workflow was configured to trigger on issue assignment events, read issue titles and bodies, and operate with read access across the organization's repositories, both public and private. For teams running agentic workflows in n8n, Claude, or similar tools, this case illustrates the core risk: any agent that reads external content and holds elevated permissions is a potential data exfiltration vector if prompt boundaries are not explicitly enforced at the system level.
OpenAI Releases GPT-5.6 Family with Three Tiers, Targets Anthropic on Coding Benchmarks
OpenAI launched GPT-5.6 on July 9, a three-variant model family consisting of Sol (flagship), Terra (mid-tier), and Luna (budget). Sol is positioned as OpenAI's strongest coding model, claiming a score of 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Anthropic's recently released Fable 5, while using less than half the output tokens and costing about one-third less. Terra scores just above Fable 5, and Luna outperforms Anthropic's Opus 4.8. OpenAI also highlights GPT-5.6 as its most capable cybersecurity model to date, supporting threat modeling, code review, patching, and blue team simulations. Sam Altman has stated Sol is 54% more token-efficient on coding tasks than prior versions. Pricing per million tokens: Sol at $5 input / $30 output, Terra at $2.50 input / $15 output, and Luna at $1 input / $6 output. The models are available now via ChatGPT, Codex, and the OpenAI API. GPT-5.6 is also the preferred model for Microsoft Copilot 365. Alongside the model launch, OpenAI introduced ChatGPT Work, an enterprise-focused tool for document drafting, spreadsheets, and presentations, available on desktop, web, and mobile. For automation builders, the new pricing tiers and coding benchmarks directly affect model selection decisions in n8n workflows and client recommendations where Claude and GPT tradeoffs are being weighed.
AgentLens Brings Trajectory-Level QA to Coding Agent Evaluation
A new benchmark called AgentLens moves beyond simple pass/fail scoring to evaluate coding agents across the full execution trajectory. The framework examines tool use, instruction-following, self-correction, and communication quality at each step an agent takes. It combines formal verification with LLM-generated trajectory reviews and side-by-side comparisons to produce structured assessments of agent behavior in production-like conditions. For teams building and QA-ing agentic workflows, the methodology offers a concrete model for evaluating agents on process quality rather than outcome alone.