The company Anthropic has announced enhancements in the AI model Claude Opus 4.6 for identifying vulnerabilities in code and shared the results of an experiment that uncovered over 500 previously unknown (0-day) vulnerabilities in the latest versions of various open-source projects. The work focused on finding vulnerabilities caused by memory management issues, as these are easier to verify. All identified vulnerabilities were assigned a high-risk level. Each vulnerability was manually checked and confirmed by Anthropic employees or external security researchers.
For the vulnerability analysis, the codebases of popular open-source projects that have been undergoing continuous fuzzing testing in the OSS-Fuzz service for a long time were utilized. Unlike traditional fuzzing testing, which generates a stream of random combinations of input data, the AI model attempted to analyze the code, taking into account past fixes to identify similar unresolved errors, highlighting problematic patterns and logically deducing which inputs might disrupt execution.
Information about the vulnerabilities identified during the experiment has already started being communicated to maintainers, with whom joint work is being conducted to implement fixes. To assist the maintainers in this manual review process, patches have been developed to address the identified issues. For example, three vulnerabilities in GhostScript, OpenSC, and CGIF have been fixed by the maintainers at the time of publication.
The configuration used to detect issues was not similar to traditional automated vulnerability detection systems; model Claude Opus 4.6 was given access to virtual machine, which, in addition to the code under investigation, included standard developer tools (coreutils, Python, etc.) and utilities for debugging and analyzing vulnerabilities (including fuzzing testing tools). The model was not provided with specific instructions on how to use these tools or special information on vulnerability detection methods. It was only given the task and the opportunity to reason about the optimal use of the available tools.
In searching for vulnerabilities in GhostScript, the AI model first attempted to conduct fuzz testing. When that did not yield results, it switched to code analysis. The code analysis also proved fruitless, prompting the model to study the git change history. In one of the commits, it noticed a mention of buffer boundary checks. Analyzing that commit, the model determined that the fix added a missing buffer boundary check when processing fonts.
Next, the model identified the code that existed before the fix and attempted to find similar usage patterns of the problematic function elsewhere in the code that remained uncorrected. As a result, a call to the gs_type1_blend function was discovered in the gdevpsfx.c file without validating the correctness of the values. Ultimately, the model compiled the contents of a file whose processing led to a crash due to data being written outside the allocated buffer.
In CGIF, the AI model was guided by the assumption that when unpacking GIF files, the library considered the size of the compressed data to always be less than that of the unpacked data. The vulnerability search focused on identifying conditions under which LZW-compressed data would exceed unpacked data sizes. Such conditions were found, and the AI model managed to create a GIT file whose processing resulted in a buffer overflow. In OpenSC, the issue was identified after analyzing the use of potentially unsafe functions strrchr and strcat in the code.
It is noted that language models have reached a level where they can identify previously unknown vulnerabilities and will soon surpass security experts in the speed and scale of vulnerability searches. It is expected that the growth in the number of identified vulnerabilities will necessitate a reform of existing disclosure processes, as the currently allotted 90-day period for fixes will be insufficient.
Source: opennet.ru
