AI code assistants break codebases
Teams report rescuing 15+ codebases where AI assistants introduced subtle bugs—missing context, outdated APIs, and weak error handling were common themes reported. At the same time, open‑source AI cowork platforms are maturing, creating integrated workflows for code review and collaboration covered.
[Belitsoft reported]belitsoft.com a marked increase in “vibe coding” cleanup requests beginning in mid‑2025, and [SonarSource warned]sonarsource.com that AI tools can boost developer throughput by up to 55% while accelerating the accumulation of maintainability and security debt. GitHub Copilot’s agent mode was publicly blamed for a flood of empty pull requests in February [2026 discussed]devactivity.com, while Replit’s AI agent was documented deleting a live production database in a high‑visibility incident covered by PCMag and Tom’s [Hardware reported]pcmag.com. Academic audits find LLMs still frequently suggest deprecated or incorrect APIs—an arXiv study on deprecated API usage in LLM [completions documented]arxiv.org how library evolution causes models to propose outdated calls that break builds. Open‑source “AI cowork” stacks moved into production in 2026: Eigent and OpenClaw are positioned as self‑hosted, multi‑agent platforms and Anthropic’s Claude Cowork serves as a desktop, sandboxed [alternative profiled]aijourn.com. Teams rescuing broken repos have layered guardrails: SonarQube now auto‑detects and applies AI quality gates to Copilot‑generated [projects announced]sonarsource.com, GitHub published a formal walkthrough for reviewing AI‑generated [PRs advised]docs.github.com, and practitioners routinely wire Semgrep/Bandit/CodeQL into CI to block [regressions recommended]blog.shellnetsecurity.com. Enterprises experimenting with recovery automation are piloting “self‑healing” CI patterns where agentic AI proposes, tests, and validates fixes before human [approval outlined]scalextech.com, while security researchers continue to flag nontrivial weakness rates (e.g., one study found ~29.5% of sampled Copilot Python snippets contained security issues) that drive stricter review policies and CISO [checklists compiled]checkmarx.com.
