What Claude's AI Code Review Actually Delivers and Where It Falls Short

While some argue about whether AI will replace junior developers, Anthropic has quietly rolled out "Code Review" for all paid teams, Cursor has ended up in the hands of SpaceX, and critical vulnerabilities with a 9.8/10 severity rating are regularly popping up in the logs of both tools. We break down point-by-point which recent news actually changes the workflow and what is just another reason for caution.
What is new with Claude Code over the last month?
Anthropic updates Claude Code almost daily — about 125 releases were pushed between the June and October versions. Notable updates: on October 14, the desktop app received nested sessions and saved environments. And on October 15, the code review tool moved out of experimental status and became generally available for organizations on Team and Enterprise plans.
Is AI code review serious or just a marketing gimmick?
The honest answer here is "both." According to Anthropic's internal data, it performs well on large changes (over 1,000 lines). That sounds impressive, but it's worth remembering:
- It does not replace a human — the company explicitly states that.
- It is less effective on small changes:
- A full review is neither free nor instantaneous — an average review takes about 20 minutes, and the feature itself is paid and depends on token volume.
The takeaway for a beginner: the tool is useful if you already have a high volume of PRs and your team physically cannot keep up with line-by-line code reading. For a solo project or a small team, it is overkill — manual review with decent linters is sufficient.
Is it safe to give an agent access to your repository?
No, and several serious incidents over the past year have proven this. Codex CLI had a vulnerability that allowed for. Cursor had a couple of critical holes found under the name DuneSlide: The attack worked without a single click —.
The practical conclusion is simple: update your tools on time, do not clone or open repositories from dubious sources in an agent, and treat the agent as an employee with broad permissions — not as a harmless autocomplete.
What are "mods" in Claude Code and should a beginner install them?
As of October 1, Claude Code introduced mods —. It sounds like a powerful customization feature, but there is a catch:. For a beginner, this means one thing — do not install the first mod you find from an unknown repository just because it promises to double your speed.
What model is actually running under the hood of Cursor?
A good reason to be skeptical of marketing. When Cursor released Composer 2, a developer discovered a model identifier in the API response via a debug proxy that revealed its origin — the Chinese Kimi K2.5 from Moonshot AI. After this was made public. By version Composer 2.5, the situation was adjusted — it used the same Kimi K2.5 base, but 85% of the compute budget was now dedicated to Cursor's own RL pipeline and post-training. The moral: not everything sold as "our own model" is built from scratch, and that is fine — but you should know exactly what you are working on.
Should a beginner chase every update?
No. With a frequency of several releases per week, it is physically impossible to test every feature manually. A reasonable tactic is to check the changelog every couple of weeks, update the version, and verify that the model's default effort settings haven't changed (as happened with the transition to Opus 5.5 in medium mode). For daily work, a stable version, clear project rules, and the habit of not trusting an agent with critical operations without review — human or instrumental — are sufficient.
In short
AI coding assistants have truly become more powerful — code review is being automated, sessions are being parallelized, and plugins are expanding capabilities. But along with power come risks: security holes, omissions about the models under the hood, and an avalanche of updates where it is easy to lose control. Treat these tools like a smart but not always honest intern — check what they are doing and do not grant unnecessary permissions.
If you want to try out modern AI assistants for code, text, and more yourself — current models are available at GEMERA AI via our Telegram bot and web dashboard, without the hassle of subscriptions and VPNs.