- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,287
It seems that the agency systems opened Pandora’s box in cybersecurity.
AI agents are already able to find old vulnerabilities cheaply and quickly. But after the report, hard work begins: check the find, repeat the failure, write a patch, release an update and bring the correct version to real systems. In one week, two news showed it at once: an autonomous agent found 21 previously unknown vulnerability in FFmpeg, and Google released Chrome 149 with a record 429 security patches.
FFmpeg is especially important here due to the scale of distribution. The library works within a huge number of programs and devices that accept, cut, recode or play videos. FFmpeg uses media services, roller processing systems, container images, Python wheels, server products, household appliances and embedded devices. If the error lives in such a code, the vulnerable component may remain in the same application for years, but immediately in dozens of dependencies, assemblies and firmware.
The depthfirst company found vulnerability. Its autonomous agent checked about 1.5 million lines of code on C and issued 21 confirmed zero-day vulnerability. For each problem, a playable example was prepared: a special file or input stream, at which the failure can be called again. According to depthfirst, the entire run cost about $1 000.
Some mistakes have been in the project for 15-20 years. Overflow of the stack in the processing code of the description table of services appeared in 2003 and lasted 23 years. For FFmpeg, this is an unpleasant fact: the project has long been used around the world, regularly tested, passed through the vasers and checked by standard means of analysis, but individual dangerous areas still remained in the code.
Most of the new finds are associated with overflows in the pile and stack. Simply put, when processing specially prepared data, the program reads or records memory outside the dedicated area. The list includes sampling formats, demultiplexers and decoders. Among the affected components are TS-duletrupexur and decoder VP9 are mentioned.
Some of the errors have already received CVE. depth numbers lists nine identifiers: from CVE-2026-39210 to CVE-2026-39218. The remaining problems, according to the researchers, are already closed in the code, but are still waiting for the official numbering. The company also posted a demonstration example so that developers and protection specialists could check their own builds.
Google released Chrome 149. In one update, the browser received 429 security patches, and for Chrome it is a record volume in one release. More than 100 vulnerabilities received critical or high levels of danger. Among the frequent classes of errors is the use of memory after release and insufficient verification of input data.
The most dangerous vulnerability from this issue was received by CVE-2026-10881 and CVSS 9.6. The error affects ANGLE, a Chrome graphics layer that helps the browser work with different graphical interfaces on different platforms. A specially prepared page can read and record data outside the permitted area of memory. With a successful attack, the attacker leaves the browser sandbox and runs the code at the operating system level. For this report, Google paid the researcher $97 000.
Record Chrome release should not be associated with one AI tool. Most of the big mistakes were found inside Google. Of the approximately 90 high-level vulnerabilities, only 10 came from external researchers. Among the 22 critical problems, 19 were found by employees. The connection with AI here is different: security teams receive more reports, quickly disassemble the same type of materials and change the requirements for the quality of applications.
In the spring, Google changed the rules of reward programs for Android and Chrome after the influx of reports prepared with AI. The company began to ask not for long descriptions, but a short exact example, which reliably reproduces the bug. The generic model easily writes a compelling text on a few pages, but an engineer needs a minimum set of actions: a file, page, or command, after which the error manifests itself without guessing.
FFmpeg has already been in similar experiments. Last year, Google Big Sleep agent reported a series of errors in the project, and some of the entries are now on the FFmpeg security page marked BIGSLEEP. The Anthropic Mythos also found bugs in this media library, including a 16-year error in H.264. According to Anthropic, several problems found by Mythos are included in the fixes of FRFmeg 8.1.
Outside the media code, the picture is similar. A few days ago, an autonomous tool found in Redis a vulnerability of remote code execution after authentication. The Bug appeared in version 7.2.0 and went unnoticed for more than two years. Another study found that the AI agent was able to prepare worker demonstrations for more than half of the 100 real vulnerabilities of the Linux kernel and bypassed the classic phases.
The phasesng checks the program with many random, strange or specially damaged input data. The method catches falls, hovering and processing errors of formats well. But some of the bugs requires an understanding of the logic of the program: which branches of the code are connected, what format will lead to the desired state, where checking data misses a dangerous case. New AI tools are trying to read large areas of the source, build hypotheses, look for a way to fail and immediately prepare an example for testing.
FFmpeg users should take a corrected build from the main project branch or security update from their distribution as soon as the patch appears. First of all, you need to check the systems that take untrusted RTSP flows or AV1 on top of the RTP. System packs check does not end. FFmpeg is often hidden inside containers, wheel packs of Python, media servers, firmware and individual video processing chains.
Chrome is also better to update without a pause. For Linux, the current build is as indicated as 149.0.7827.53. For Windows and macOS, versions 149.0.7827.53 and 149.0.7827.54 are available. With the updated on, you need to check that the browser has already been restarted and the new version is installed.
The search for bugs fell sharply. The correction did not fall in price with him. The report needs to be read, the failure to reproduce, the reason to find, write a patch, drive tests, release the assembly, update the dependence, reassemble the container, bring the device to a new firmware. Most of this work remains on accompanying open projects, internal development teams and several people who now have to disassemble machine search results.
The old pace of updates can no longer withstand such a load. Correction of dependencies with CVE cannot be postponed. Short patch cycles, auto-renewal and rapid reassembly of components become the basic protection: the car can be started for $ 1 000 and get dozens of old vulnerabilities in the library, which for years lay almost everyone at hand.
AI agents are already able to find old vulnerabilities cheaply and quickly. But after the report, hard work begins: check the find, repeat the failure, write a patch, release an update and bring the correct version to real systems. In one week, two news showed it at once: an autonomous agent found 21 previously unknown vulnerability in FFmpeg, and Google released Chrome 149 with a record 429 security patches.
FFmpeg is especially important here due to the scale of distribution. The library works within a huge number of programs and devices that accept, cut, recode or play videos. FFmpeg uses media services, roller processing systems, container images, Python wheels, server products, household appliances and embedded devices. If the error lives in such a code, the vulnerable component may remain in the same application for years, but immediately in dozens of dependencies, assemblies and firmware.
The depthfirst company found vulnerability. Its autonomous agent checked about 1.5 million lines of code on C and issued 21 confirmed zero-day vulnerability. For each problem, a playable example was prepared: a special file or input stream, at which the failure can be called again. According to depthfirst, the entire run cost about $1 000.
Some mistakes have been in the project for 15-20 years. Overflow of the stack in the processing code of the description table of services appeared in 2003 and lasted 23 years. For FFmpeg, this is an unpleasant fact: the project has long been used around the world, regularly tested, passed through the vasers and checked by standard means of analysis, but individual dangerous areas still remained in the code.
Most of the new finds are associated with overflows in the pile and stack. Simply put, when processing specially prepared data, the program reads or records memory outside the dedicated area. The list includes sampling formats, demultiplexers and decoders. Among the affected components are TS-duletrupexur and decoder VP9 are mentioned.
Some of the errors have already received CVE. depth numbers lists nine identifiers: from CVE-2026-39210 to CVE-2026-39218. The remaining problems, according to the researchers, are already closed in the code, but are still waiting for the official numbering. The company also posted a demonstration example so that developers and protection specialists could check their own builds.
Google released Chrome 149. In one update, the browser received 429 security patches, and for Chrome it is a record volume in one release. More than 100 vulnerabilities received critical or high levels of danger. Among the frequent classes of errors is the use of memory after release and insufficient verification of input data.
The most dangerous vulnerability from this issue was received by CVE-2026-10881 and CVSS 9.6. The error affects ANGLE, a Chrome graphics layer that helps the browser work with different graphical interfaces on different platforms. A specially prepared page can read and record data outside the permitted area of memory. With a successful attack, the attacker leaves the browser sandbox and runs the code at the operating system level. For this report, Google paid the researcher $97 000.
Record Chrome release should not be associated with one AI tool. Most of the big mistakes were found inside Google. Of the approximately 90 high-level vulnerabilities, only 10 came from external researchers. Among the 22 critical problems, 19 were found by employees. The connection with AI here is different: security teams receive more reports, quickly disassemble the same type of materials and change the requirements for the quality of applications.
In the spring, Google changed the rules of reward programs for Android and Chrome after the influx of reports prepared with AI. The company began to ask not for long descriptions, but a short exact example, which reliably reproduces the bug. The generic model easily writes a compelling text on a few pages, but an engineer needs a minimum set of actions: a file, page, or command, after which the error manifests itself without guessing.
FFmpeg has already been in similar experiments. Last year, Google Big Sleep agent reported a series of errors in the project, and some of the entries are now on the FFmpeg security page marked BIGSLEEP. The Anthropic Mythos also found bugs in this media library, including a 16-year error in H.264. According to Anthropic, several problems found by Mythos are included in the fixes of FRFmeg 8.1.
Outside the media code, the picture is similar. A few days ago, an autonomous tool found in Redis a vulnerability of remote code execution after authentication. The Bug appeared in version 7.2.0 and went unnoticed for more than two years. Another study found that the AI agent was able to prepare worker demonstrations for more than half of the 100 real vulnerabilities of the Linux kernel and bypassed the classic phases.
The phasesng checks the program with many random, strange or specially damaged input data. The method catches falls, hovering and processing errors of formats well. But some of the bugs requires an understanding of the logic of the program: which branches of the code are connected, what format will lead to the desired state, where checking data misses a dangerous case. New AI tools are trying to read large areas of the source, build hypotheses, look for a way to fail and immediately prepare an example for testing.
FFmpeg users should take a corrected build from the main project branch or security update from their distribution as soon as the patch appears. First of all, you need to check the systems that take untrusted RTSP flows or AV1 on top of the RTP. System packs check does not end. FFmpeg is often hidden inside containers, wheel packs of Python, media servers, firmware and individual video processing chains.
Chrome is also better to update without a pause. For Linux, the current build is as indicated as 149.0.7827.53. For Windows and macOS, versions 149.0.7827.53 and 149.0.7827.54 are available. With the updated on, you need to check that the browser has already been restarted and the new version is installed.
The search for bugs fell sharply. The correction did not fall in price with him. The report needs to be read, the failure to reproduce, the reason to find, write a patch, drive tests, release the assembly, update the dependence, reassemble the container, bring the device to a new firmware. Most of this work remains on accompanying open projects, internal development teams and several people who now have to disassemble machine search results.
The old pace of updates can no longer withstand such a load. Correction of dependencies with CVE cannot be postponed. Short patch cycles, auto-renewal and rapid reassembly of components become the basic protection: the car can be started for $ 1 000 and get dozens of old vulnerabilities in the library, which for years lay almost everyone at hand.
