Skip to content

Topic

AI-assisted vulnerability research

Use of frontier and open-weight models to discover, triage and remediate software vulnerabilities, and the resulting shift in attacker-defender economics.

Current stories

securityConfirmed4 publishers

XBOW built a Windows Chrome exploit chain in six days with about $40,000 of AI compute

XBOW chained three unknown Chrome bugs into a Windows exploit that runs on page load, built in six days with a public OpenAI model, it said. Google fixed all three and paid its $250,000 top reward; the fixes apply only after Chrome restarts.

Publishers:cryptobriefing.com dev.to thehackernews.com xbow.com

Perspective Coverage

4 publishers
Builder
Builder 25%
Operator
Operator 41%
Investor
Investor 34%

Reality

Evidence62
Adoption
Not scored
Hype gap+30
Incentives70
Confidence72
securityOne report1 publisher

Guest VMs can read host memory through Linux kernel flaw PwnEscape

Linux kernel flaw CVE-2026-99999 let pwn.ai's Paulos Yibelo read about 32 bytes of host memory per packet from Vercel's production sandboxes. Operators who host other people's VMs need the fix in the kernel the host itself runs.

Reality

Evidence66
Adoption
Not scored
Hype gap+8
Incentives50
Confidence62
buildOne report1 publisher

Windows Shell zero-day CVE-2026-96951 strips the tag that SmartScreen and Protected View read

Microsoft's October Patch Tuesday fixes CVE-2026-96951, an exploited Windows Shell zero-day rated CVSS 8.8 that strips Mark-of-the-Web from files. A dev.to analysis counts it as the third such bypass this year.

Publishers:dev.to

Reality

Evidence40
Adoption
Not scored
Hype gap+15
Incentives
Not scored
Confidence45
buildConfirmed5 publishers

OpenSSH 10.6 turns off LZ77 compression to stop a cross-channel plaintext leak

OpenSSH 10.6, released October 6, turns off the LZ77 coder in SSH compression to block an attack that reads one channel's secrets from another. Hosts using the Compression option will get less from it, so the project suggests compressing in the application.

Perspective Coverage

5 publishers
Builder
Builder 42%
Operator
Operator 52%
Investor
Investor 6%

Reality

Evidence82
Adoption
Not scored
Hype gap+5
Incentives20
Confidence78
securityOne report1 publisher

Attackers scan Rejetto HFS servers for a signing-key flaw that grants admin control

VulnCheck says attackers are probing internet-facing Rejetto HFS servers for CVE-2026-61500, a signing-key flaw that gives full administrative control. The bug affects HFS 3.0.0 through 3.2.0, and the fix shipped in 3.2.1.

Reality

Evidence55
Adoption
Not scored
Hype gap+30
Incentives40
Confidence60
productOne report1 publisher

Apple patches a Screen Sharing flaw that hands network attackers root on an unpatched Mac

Apple's emergency macOS update on August 6 fixed a Screen Sharing flaw that calif.io turned into a working exploit about four hours later. The flaw let anyone on the network sign in without a password, and the service answering those connections runs as root.

Publishers:blog.calif.io

Reality

Evidence55
Adoption
Not scored
Hype gap+10
Incentives45
Confidence50
buildOne report1 publisher

HFS leaks the Math.random() state that signs its admin cookies

Rejetto HFS 3.0.0 through 3.2.0 signs session cookies with a Math.random() key, and five leaked PRNG outputs let an attacker forge an admin session. Exploitation began on October 1, and version 3.2.1 restores secure key generation.

Publishers:dev.to

Reality

Evidence70
Adoption72
Hype gap0
Incentives50
Confidence65
securityOne report1 publisher

Anthropic's Mythos model found a session-forgery path to RCE in Rejetto HFS 3.X

Horizon3 used Anthropic's Mythos model to find CVE-2026-61500, a chain that forges Rejetto HFS admin sessions and reaches remote code execution. The firm expects frontier models to make deeper, less reliable bug classes worth weaponizing at scale.

Publishers:horizon3.ai

Reality

Evidence55
Adoption15
Hype gap+30
Incentives70
Confidence50
buildOne report1 publisher

GitHub Security Lab credits step-by-step LLM taskflows with 24 Android vulnerability reports

GitHub Security Lab says its open-source LLM taskflows have found and reported 24 Android app vulnerabilities by auditing code in staged steps. Teams with a GitHub Copilot license can point the same audit at their own repositories.

Reality

Evidence45
Adoption
Not scored
Hype gap+20
Incentives60
Confidence55
securityOne report1 publisher

GitHub Security Lab's AI taskflows turned up 24 flaws in Android apps including OsmAnd and Wikipedia

GitHub Security Lab's Kevin Stubbings used AI audit workflows to find and report 24 Android app flaws, among them bugs in OsmAnd and Wikipedia. The workflows are free to run on any repository, so the same audit is open to defenders and attackers alike.

Reality

Evidence50
Adoption
Not scored
Hype gap+20
Incentives45
Confidence45
securityConfirmed8 publishers

Exposed MikroTik SSH hands over full administrative control without authentication

CERT Polska dated successful attacks to at least September 2 and published its warning on September 5, so operators who deferred the RouterOS update have three days of configuration changes to read as well as a patch to install.

Perspective Coverage

8 publishers
Builder
Builder 19%
Operator
Operator 73%
Investor
Investor 8%

Reality

Evidence78
Adoption45
Hype gap+18
Incentives30
Confidence72
securityConfirmed5 publishers

Researchers built a WeChat worm that hijacks accounts while the phone is still ringing

Calif's WeWorm abused a memory corruption bug in WeChat's VoIP stack to take over accounts on an iPhone 17e and two Pixel 10a handsets, and Tencent blocked it server-side on 28 August without publishing an advisory or a CVE.

Perspective Coverage

6 publishers
Builder
Builder 32%
Operator
Operator 51%
Investor
Investor 17%

Reality

Evidence55
Adoption
Not scored
Hype gap+35
Incentives60
Confidence60
securityOne report1 publisher

Disabling unprivileged user namespaces stops three of four new Linux local-root bugs

Fixes are in for four Linux local-root bugs found in code more than a decade old. Turning off unprivileged user namespaces closes three of them. The fourth stays open.

Publishers:heyitsas.im

Reality

Evidence58
Adoption32
Hype gap+12
Incentives55
Confidence55
securityOne report1 publisher

Asim Manizada published working local-root exploits for four freshly patched Linux kernel bugs

The fixes are already upstream, so the work now is confirming your distribution shipped them before someone with a low-privileged shell on a shared host uses the published code to reach root.

Reality

Evidence62
Adoption35
Hype gap−10
Incentives35
Confidence60
securityOne report1 publisher

Chainguard opens Athena's 40,000-finding backlog on September 28 with fifty silent fixes

Chainguard says fix generation was never its bottleneck and that responsible disclosure at scale is, so its first public batch is built from bugs upstreams quietly fixed years ago and never filed CVEs for.

Publishers:chainguard.dev

Reality

Evidence34
Adoption16
Hype gap+28
Incentives82
Confidence48
scienceOne report1 publisher

Stripping refusals from Gemma tilted its bug verdicts toward yes

A researcher ran the same FreeBSD scan through base and abliterated open-weight builds and found the uncensored ones graduating three to four times as many findings, while the most aggressive one never surfaced the actual CVE.

Publishers:clearbluejar.github.io

Reality

Evidence58
Adoption14
Hype gap+14
Incentives22
Confidence46
buildOne report1 publisher

Disproving one pointer in Lemmalog retracts every conclusion that rested on it

Jordy Zomer built a Datalog engine so an agent maintains what it currently knows instead of searching its own transcript, and his own benchmark runs put the weak link in the model that writes the facts.

Publishers:runtimewire.com

Reality

Evidence46
Adoption12
Hype gap−8
Incentives28
Confidence54
securityOne report1 publisher

Kaspersky's CVE surge has two sources, and only one of them lands in the CVE count

Kaspersky's Q2 2026 figures credit AI with both writing the bugs and finding them. The disclosures carrying working exploit code never got a CVE identifier at all.

Publishers:securelist.com

Reality

Evidence48
Adoption44
Hype gap+30
Incentives66
Confidence45

Earlier coverage

  1. Grayscale's ZCSH puts Zcash in brokerage accounts, and none of its ZEC is shielded

    Invest · August 25, 2026 · Confirmed3 publishers