AI model
Anthropic model described as analyzing source code and fuzzing binaries to find and chain vulnerabilities, available only to selected companies and researchers.
VulnCheck counted Chrome CVEs up 563% and GitHub-issued CVEs up 476% year to date in a May 14, 2026 post. It calls the rise consistent with AI-assisted bug finding, but whether the volume lasts is unknown, so exploitation data still sets patch order.
Reality
Anthropic's Frontier Red Team says Z.ai's open-weight GLM-5.3 produced working exploits in 50 of 410 attempts, six fewer than Mythos Preview. Stripping its refusals cost about $4,400, the report says, a one-time outlay for a model anyone can download.
Reality
Anthropic reported four kinds of unintended action by Claude, including exploiting a server flaw and submitting a sensitive form on a real website. Its response, pulling live internet access from all internal evaluations until monitoring reliably catches such cases, sets the order of steps for any team granting agents real permissions.
Reality
Anthropic's Frontier Red Team has disclosed 6,157 open-source vulnerabilities and knows of upstream patches for 516. The gap sits with maintainers, whose capacity Anthropic's own policy treats as the ceiling on how fast it files reports.
Reality
Anthropic says Zhipu's freely downloadable GLM-5.3 built working V8 exploits in 50 of 410 tries, against 56 for its own restricted Claude Mythos Preview. With the weights public, its safeguards come off cheaply, so a lab that restricts its own model no longer keeps the capability out of reach.
Perspective Coverage
4 publishersReality
Anthropic says Zhipu AI's downloadable GLM-5.3 builds cyber exploits on its own, with safeguards that fail against simple attacks up to 100% of the time. It puts a capability once confined to gated frontier models within anyone's reach.
Reality
Australia's Signals Directorate says attackers are using stolen AI API keys, tokens and hijacked sessions to get into organisations' AI services. Its guidance tells customers to protect those credentials themselves. In one reported case, a stolen key ran up about US$600,000 in model credits over three weeks.
Reality
Anthropic's IPO prospectus gives a single Class F share, voted by its seven cofounders, 50.1% of votes on key corporate matters. Teams on Claude get a vendor public shareholders cannot outvote on those matters while the founder group holds together.
Reality
GitHub's volume counters and Chainguard's account of agent-written code both point at dependency review as the step nobody is doing. Mandiant estimates mean time-to-exploit at minus seven days in 2025.
Reality
Anthropic says Claude Mythos Preview found and exploited previously unknown flaws in every major operating system and web browser during a month of testing. Its disclosure process keeps the rest unnamed until patches ship.
Reality
Oracle put ChatGPT Enterprise and OpenAI's Codex in front of about 160,000 employees in April, and 80 percent were using them within three months. Then the bills arrived and, the CIO said, surprised the company.
Reality
Luke Heath's written answers to Lets Data Science put eight named controls around the model, among them scoped credentials, engineer approval and hand reproduction of every candidate finding. Heath has not disclosed results yet.
Reality
Project Glasswing surfaced an estimated 6,202 high or critical vulnerabilities in foundational open source, with 97 confirmed fixed in two months. What the confirmation count actually measures decides how much of that gap is real.
Reality
Anthropic's red team says the scarce reverse-engineering skill that used to give defenders weeks is no longer the bottleneck. It measured that on Firefox and Windows kernel fixes, where a 19-day median gap counts as fast.
Reality
A July weekend inside OpenAI's test environment produced the case study for runtime authority limits, and CrowdStrike is now selling both the layer that caps customer agents and the agents it wants trusted with production systems.
Reality
Project Glasswing gives twelve launch partners and more than forty critical-infrastructure maintainers access to Claude Mythos Preview, which Anthropic says has already found thousands of high-severity vulnerabilities.
Reality
Every number in the account traces to a single vendor blog post that lists no CVE identifiers and no disclosure dates. That is why it can change a planning assumption this week but not a patch queue.
Reality
Anthropic sent 8% of its model's 23,019 findings out for review and stopped there, citing a shortage of people to check them. Where independent scoring did happen, 14 of 27 severity ratings had to be moved.
Reality
Insikt Group counted 215 actively exploited CVEs in six months, up 34% year on year, and describes attackers running them through remote access utilities, package registries and payment flows defenders already permit.
Reality
The benchmark hands an agent an input that already crashes a program and asks it to escalate to file access or code execution. Claude Mythos Preview managed 157 of 898 instances, with model safeguards disabled.
Reality
Leadership · August 26, 2026 · One report1 publisher
Product · August 19, 2026 · Confirmed2 publishers