NebuSec
As a security research group that has spent months building and deploying AI agents for vulnerability research, our perspective is grounded in operational reality: real-world bug discovery, exploit generation, and coordinated disclosure.
Our AI agent has discovered over 300 bugs in the Linux kernel, and several high-risk zero-day vulnerabilities in Google Chrome V8 JS engine, leading to multiple wins in kernelCTF, and over $100,000 in bug bounty. Our AI agent has also found bugs in WordPress, CPython, OpenSSL, and macOS.
All of these findings were carefully validated and responsibly disclosed, accompanied by detailed root cause analyses and minimal reproducers, and were acknowledged by both the Linux kernel community and Google. We are also making every effort to help get these bugs fixed by developing and rigorously vetting patch proposals ourselves before sending them to maintainers. Even with LLM assistance, every fix still requires careful human review to ensure that it is both correct and well designed.
Find more exploitable vulnerabilities over lunch without relying on Mythos.
Has Mythos fundamentally reshaped cybersecurity? No
No. LLMs had already changed the field well before Mythos. What Anthropic presents is not a fundamental shift, but mostly incremental progress—much of it likely driven by engineering rather than by a uniquely capable private model. The findings Anthropic highlights also appear to include relatively few truly high-impact vulnerabilities, while its clearest public exploitation example involves an unusually weak legacy target. By combining a small number of respectable but fairly ordinary findings with a much larger set of bugs that seem to have limited real-world exploit value, Anthropic creates an inflated impression of what Mythos has actually accomplished. The result is a public narrative that can easily be read as “the model found a thousand zero-days,” even though the actual security significance appears considerably narrower.
Can "inferior" publicly available models achieve the same results or even do better? Yes
Yes. Using GPT-5.4 and GPT-5.3, our AI agent identified more than 300 bugs in three days at a cost of under $20k (while generating 5x return in bug bounty). By contrast, Anthropic has publicly described only four distinct browser exploits and three Linux kernel exploits. Our own pipeline has already produced six browser exploits across six separate zero-days, along with three Google COS kernel exploits, with 21 additional exploitable cases still in progress. More importantly, we focus on real security impact, not inflated totals that count low-severity crashes or denial-of-service findings as if they were major zero-days.
Do human researchers still matter? Absolutely yes!
In fact, the Mythos story helps illustrate why. With the right expertise, system design, and harnessing, today’s general-purpose models are already capable of matching or even exceeding supposedly more advanced private models on broad security tasks. Anthropic relied on engineering as well. The difference is that its public narrative presents much of that system-level gain as though it were primarily a model breakthrough.
Bugs List
Project Description ID Type
- Google V8 Turbofan: [REDACTED] Issue 497404188 Type Confusion
- Google V8 Maglev: [REDACTED] Issue 494914816 Type Confusion
- Google V8 Maglev: [REDACTED] Issue 493534950 Type Confusion
- Google V8 Maglev: [REDACTED] Issue 497112471 Type Confusion
- Google V8 Maglev: [REDACTED] Issue 493951261 Type Confusion
- Google V8 Maglev: incorrect phi untagging may lead to exploitable write barrier omission CVE-2026-5865 Type Confusion Exploitable arbitrary memory read/write. Chrome 146 and earlier versions, Windows, Linux and macOS
- Linux kernel REDACTED: Remote out-of-bounds write may lead to privilege escalation via limited memory write kernelCTF exp453 Out of Bounds Affects Google COS, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43 (most devices with IPv6 enabled are affected)
- Linux kernel REDACTED: A generic Use-after-free may lead to privilege escalation via arbitrary memory write [REDACTED] Use After Free Affects nearly ALL mainstream Linux distributions, including: Android 16, Google COS, Ubuntu 24, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Out-of-bounds write may lead to privilege escalation via arbitrary memory read/write kernelCTF exp464 Out of Bounds Affects Google COS, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel netfilter: uninitialized memory use may lead to privilege escalation via control-flow-hijack CVE-2026-23274 (kernelCTF exp457) Use Before Initialization Affects Google COS, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to to privilege escalation via control-flow-hijack Report Hash: b70c4ab334c2bb8dc6c2af8e17f54fa4 Use After Free Affects Google COS, Ubuntu 24, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via arbitrary memory read/write Report Hash: cd4f78294a76b3e9cdbe5ee60b5fe274 Use After Free Affects Google COS, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via control-flow-hijack Report Hash: 09f00f380a864bc9863b5841b51caeef Use After Free Affects Google COS, Ubuntu 24, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via control-flow-hijack Report Hash: 3dc44256f1648df1d9e84f33006aa978 Use After Free Only affects Debian 12
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via control-flow-hijack Report Hash: 665ff5b1e7f08cb1b30f356d9bd8f76e Use After Free Only affects Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via control-flow-hijack and/or arbitrary memory read/write Report Hash: 2b2b37f9d9082fe4718c2dfee94c6a9a Use After Free Affects Debian 12, RedHat 10.1 and CentOS 9
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via control-flow-hijack Report Hash: 0a44fc5252db302242f28e0c5a970bee Use After Free Affects Debian 12, RedHat 10.1 and CentOS 9
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via control-flow-hijack Report Hash: 57dc56430cd217ed292a9c814e5b3e87 Use After Free Affects RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via arbitrary memory read/write Report Hash: f6ce8deafe8cbb8d1bc7f33e3db43645 Use After Free Affects Debian 12, RedHat 10.1 and CentOS 9
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via arbitrary memory read/write Report Hash: de110b48d8afd98e1960f65a2f2828b1 Use After Free Affects Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via arbitrary memory read/write Report Hash: 3baa69563168807ccb52cae21b2b3fcc Use After Free Only affects Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via arbitrary memory read/write Report Hash: 8e53b507e533eeece29d67a393b77351 Out of Bounds Affects Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via arbitrary memory read/write Report Hash: 98236c00d46556e5d9b95ff9d9a8e24f Use After Free Affects Fedora 43 and Arch Linux
- Linux kernel REDACTED: Invalid-free may lead to privilege escalation via arbitrary memory read/write Report Hash: db06e0ea08496808677689cb88333783 Invalid Free Affects Debian 12, Ubuntu 24 and Fedora 43
- Linux kernel REDACTED: Double-free may lead to privilege escalation via arbitrary memory read/write Report Hash: 684317a0b9c853c6cc0624b8f95dfcf92 Double Free Affects Fedora 43 and Arch Linux
- Linux kernel REDACTED: Double-free may lead to privilege escalation via arbitrary memory read/write Report Hash: cd4f78294a76b3e9cdbe5ee60b5fe274 Double Free Affects Google COS, Debian 12, RedHat 10.1, CentOS 9 and Fedora 43
- Linux kernel REDACTED: Use-after-free may lead to privilege escalation via control-flow-hijack Report Hash: 4d15701c07ab3cc84fbd94749b3e41b1 Use After Free Affects Fedora 43 and Arch Linux
- Linux kernel And 67 additional unprivileged-user-triggerable memory corruption bugs > read more Report Hashes in Reference link Memory Corruption
- Linux kernel And 171 more additional default-root-triggerable memory corruption bugs > read more Report Hashes in Reference link Memory Corruption
- Linux kernel And 129 additional trivial DoS bugs, including unprivileged-user-triggerable and default-root-triggerable cases > read more Report Hashes in Reference link Denial of Service
- WordPress Authorization bypass in [REDACTED] CVE-2026-23765 [RESERVED] Authorization Bypass
- WordPress Insufficient cryptographic validation may affect signature verification CVE-2026-25074 [RESERVED] Signature Forgery
- WordPress Stored XSS via [REDACTED] CVE-2026-28533 [RESERVED] Stored XSS
- WordPress Stored XSS via [REDACTED] CVE-2026-28534 [RESERVED] Stored XSS
- WordPress Stored XSS via [REDACTED] CVE-2026-28998 [RESERVED] Stored XSS
- WordPress Stored XSS via [REDACTED] CVE-2026-28999 [RESERVED] Stored XSS
- CPython Improper ZIP extraction handling may allow writing files outside the target directory CVE-2026-3087 [RESERVED] Path Traversal
- CPython Remote out-of-bounds write may lead to denial of service and potentially remote code execution CVE-2026-3298 [RESERVED] Out of Bounds