Brief IA
›

CodeRabbit Claims 53,000 Critical Reports in Ten Days

AI Tools·Tom Levy·

CodeRabbit Claims 53,000 Critical Reports in Ten Days

CodeRabbit Claims 53,000 Critical Reports in Ten Days
Key Takeaways
1CodeRabbit reports approximately 53,000 High or Critical alerts over ten days in early November.
2Microsoft reports that an agent extracted Azure Key Vault data in five minutes after searching for GitHub credentials.
3The estimate of $90.20 put forward by Tomasz Tunguz pertains solely to the use of the model.
Why it matters : The controls proposed by CodeRabbit also cover legacy code, beyond just the changes submitted in PR.

CodeRabbit claims approximately 53,000 reports of High or Critical severity following a scanning campaign involving around 2,000 organizations. Its tools combine change review, repository exploration, and patch proposals, while Microsoft and Anthropic describe offensive operations conducted by agents.

CodeRabbit announces 1,049 high-severity fixes

Over ten days in early November, CodeRabbit claims to have identified around 53,000 results classified as High or Critical across approximately 2,000 organizations using its Deep Scans. The company also reports over 10,000 results corrected, of which 1,049 fall into these severity categories. It has committed up to $1 million in free scans aimed at researching and fixing hidden vulnerabilities.

To address the reports, CodeRabbit offers CodeRabbit Agent, which generates reviewable patches. At a developer's request, this agent can apply a fix and then open a PR, with the ability to inspect the vulnerability, the proposed change, and the associated evidence. According to CodeRabbit, configurable severity thresholds also allow for the automatic creation of tickets in Jira or Linear to integrate remediation into existing tracking.

Microsoft describes a breach achieved in five minutes

Microsoft reports that an agent took five minutes to move from searching for exposed credentials on GitHub to extracting data from Azure Key Vault. When its cloud queries failed, it adjusted API versions and renewed its attempts until successful. Anthropic describes campaigns conducted with agent frameworks against multiple victims: reconnaissance, exploitation, and extraction were executed simultaneously on different targets. According to Anthropic, these operations, previously reserved for specialized teams, required less manual intervention.

The cost of such attacks must be distinguished between model consumption and total expenses. Tomasz Tunguz estimates the use of GLM-5.3 at $90.20 for an offensive load of 100 million tokens, priced at $0.902 per million. This estimate is presented as being approximately 98% to 99% lower than the $5,000 to $10,000 cited by the Bank for International Settlements for a comparable cyberattack. However, Tunguz's calculation only covers model usage, excluding other costs of an attack.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Pillar exposes a flaw in OpenCode authentication

Tools connected to agents can serve as additional entry points. In October, Pillar disclosed a vulnerability concerning OpenCode: according to the company, a malicious Model Context Protocol (MCP) server could execute commands during authentication on Windows or WSL. The URL sent by this server reached a shell without needing to deceive the model. Also in October, OpenAI models contributed to the discovery of two PHP2Shell vulnerabilities, which allow for remote code execution in the PHP runtime.

The risk is not limited to known vulnerabilities: a deployed application could also be attacked by bypassing a permission or business rule. CodeRabbit presents Secure's semantic analysis as a means to track requests in the code and verify these controls. The company claims that its agents examine attack paths and validate their conclusions independently of the person or agent behind the code.

For incoming changes, CodeRabbit announces a review of each commit associated with a new PR, including when the code comes from an agent. The review targets vulnerabilities that are actually accessible and exploitable in the affected application. The company also indicates that it studies the effect of architectural changes on permissions, the flow of sensitive data, and trust boundaries.

Deep Scan also tracks unexamined code portions

At the repository level, Deep Scan also covers older code paths not affected by the latest PR, according to CodeRabbit. The agents catalog entry points, sensitive operations, and access controls, then continue their investigations between functions and files. When configuration allows, associated repositories add context from shared libraries and services. A separate agent must challenge each presumed vulnerability by verifying its scope, exploitation conditions, and protections that could invalidate it.

Each review produces, according to the company, a mapping distinguishing controlled entries, those never studied, and those whose verification has become obsolete after code modification. Delta scans revisit modified files as well as areas that are unverified or have become outdated. This tracking informs teams which part of the repository has been examined for security and when that examination took place. The offering also includes validation of results during execution.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.