Technical Context
I wouldn't call this a sensation, but the pattern is too recognizable: people are reporting that Claude Code has started defaulting to a defensive mode, flagging prompt injections even on harmless tasks. For those building AI integration into their development process, this isn't a minor glitch—it's a direct hit to pipeline predictability.
In parallel, a very down-to-earth workaround has surfaced: the OpenAI Codex Plugin for Claude Code. Discussions frequently mention the commands /codex:rescue and /codex:adversarial-review, along with advice to update Codex to the latest version and set xhigh reasoning effort. I appreciate setups like this not for their magic, but because they transform a single, temperamental agent into a system with a backup circuit.
The idea itself is simple and powerful: instead of trying to coax one LLM into being a generator, a validator, and a paranoiac all at once, you separate the roles. Claude writes the code, and Codex attacks it as a critic, searching for edge cases, vulnerable assumptions, and logical holes. I particularly liked one technique: telling Claude upfront that its code will be reviewed by Codex. This noticeably changes the output style, as the model cuts fewer corners.
The most notable observation from the threads, which I'd treat as a user case rather than a scientific benchmark, is this: one person ran over 280 experiments overnight on a 20x subscription and saw about a 10% quality improvement while they slept. I wouldn't take the numbers as absolute, but the principle is familiar: adversarial critique almost always catches what a single prompt misses.
Impact on Business and Automation
The winners here are teams that have already integrated code generation into their process, rather than using it as a toy. If one agent becomes unstable, a second verification layer saves deadlines, nerves, and iteration costs. This is often cheaper and faster than endlessly re-prompting Claude, hoping it will correct itself this time.
The losers are those building an AI architecture based on a 'one model solves all' scheme. In practice, a combination of roles works more reliably: generation, critique, a rescue scenario, and clear escalation rules for when an agent starts to panic or argue with reality.
At Nahornyi AI Lab, we solve these kinds of issues for clients regularly. We don't just plug in a model; we build a functional AI automation system with checks, fallback logic, and a manageable cost of error. If your code agents are already slowing down your team, let's analyze your workflow and build an AI solution that delivers results overnight, not new surprises by morning.