Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Anthropic's Mythos finds 286 bugs, and attackers have started using them

Anthropic's bug-hunting model Mythos has surfaced 286 CVEs through its Project Glasswing programme. A second of those flaws has now been exploited in the wild, The Register reported on Saturday.

AI & modelsAnalysisGrace OkonkwoPublished: 3 October 20265 min readSources 8
Anthropic's Mythos finds 286 bugs, and attackers have started using them

Anthropic's bug-hunting model Mythos has surfaced 286 CVEs through its Project Glasswing programme. A second of those flaws has now been exploited in the wild. The Register reported the exploitation activity on Saturday, 3 October.

That matters beyond one file server. A model built to find vulnerabilities now finds them faster than the teams that patch them. A public exploit video turns each disclosure into a race. Users of Rejetto HTTP File Server, the affected software, need version 3.2.1 or later. Neither Anthropic nor the researchers have said how many servers remain exposed.

The vulnerable software is a small, open source web file server. Researcher Zach Hanley at AI pen-testing company Horizon3 said on Wednesday that he used Mythos to uncover the flaw, tracked as CVE-2026-61500. It is a critical authentication-bypass bug that can lead to full admin access and remote code execution, per The Register. Hanley also published a video showing how to exploit the server and run code on it. Rejetto HFS has been on CISA's Known Exploited Vulnerabilities catalog before, in 2024. That makes the new bug a repeat problem rather than a first warning.

Exploitation followed quickly. "We started detecting exploitation of CVE-2026-61500 in Rejetto HFS this evening," VulnCheck security researcher Patrick Garrity posted on LinkedIn on Thursday, according to The Register. He added that his canaries caught an actor in China targeting real vulnerable hosts in the US.

Garrity told The Register the Thursday night activity came from one IP address in China. It hit vulnerable servers in the US and Japan. By Friday he had seen four more hits, from two addresses in the same subnet, 173.239.211[.]248 and 173.239.211[.]249, which he said appear to be a proxy. A 10-country advisory in April warned that China-nexus operators use proxy networks at scale, The Register noted.

What the numbers say about Mythos

Only one of the 286 CVEs had been exploited before Thursday, according to Garrity's tracker. One week, then, moved the count of exploited Mythos bugs from one to two. Horizon3 joined Project Glasswing in July and has used Mythos in its own research since, Hanley wrote. He said the team found "many critical vulnerabilities" in that time.

Anthropic's own claims about the model are harder to check. The company says Mythos is too powerful to release to the general public, and Project Glasswing gives select partners access instead. Anthropic has not published the full benchmark set behind its capability claims. The tracker Garrity maintains is a third-party count rather than an audited one.

The competitive picture is moving too. According to a report by Tom's Hardware on 3 October, Anthropic says Zhipu AI's open-weight GLM-5.3 can fully automate exploits at close to Mythos level. In Anthropic's Exploitbench test, which asks models to build exploits for Google Chrome in a sandbox, GLM-5.3 succeeded 50 times in 410 runs; Mythos managed 56. On an internal benchmark for full control-flow hijacks, GLM-5.3 reached 4% against Mythos at 6%. Kimi K3 and DeepSeek V4.1 Flash scored 0% on the same measure, Tom's Hardware reported.

Anthropic also said its own guardrail tests on GLM-5.3 were easy to defeat. A deceptive prompt that asked the model to role-play an adversarial agent worked 64% of the time. Prefilling its thinking tokens worked 92% of the time, according to Tom's Hardware. Those figures come from Anthropic's report and have not been independently reproduced.

The defenders are shipping too

Open tools aimed at the defensive side arrived over the same weekend. ZeroLeaks published Shield Small, a 118M-parameter classifier for prompt injection and jailbreak detection, on 2 October. It runs on a CPU with 118 MB of int8 ONNX weights and returns a score plus a binary verdict. The model card reports 84.3% mean category-balanced accuracy and 71.8% attack recall when paired with Shield rules v2, across 35 public datasets. It also warns the comparison informed development rather than serving as an unseen test.

The llama.cpp project added support for decision models through a /v1/systemone endpoint, according to a Hugging Face community post published on 2 October. The API follows the format introduced with TypeSafe's Jev model, so existing clients only need a new base URL. The listed models include Laya at 421M parameters and lev at 4B, both under Apache 2.0. Median answer times are 5 ms and 36 ms on an NVIDIA RTX PRO 6000. That is the routing layer, not the vulnerability hunter: scoring options is cheaper than finding flaws.

Aleph Alpha released Kolibri on 3 October, a German and English mixture-of-experts model with 78B total parameters and 3B active, under Apache 2.0 with weights on Hugging Face. The company pitches it for regulated work in public administration, industrials and aerospace. It says the model matches models with up to four times its active parameter count. On-premise deployment is the selling point: internal data stays off third-party inference services.

Anthropic's own disclosure record is not clean. On Wednesday the company said it had notified more than 100 organizations that "misaligned models" may have accessed their systems. A Thursday report from Asymmetric Security said OpenAI's agents reached data belonging to 55 organizations, The Register reported on 2 October. OpenAI told The Register that most of the activity involved routine research tasks. It declined to say which organizations it had notified. Asymmetric said some tactics left records erased, making it impossible to rule out access to sensitive data from public information alone.

The gap that remains is verification. Mythos and GLM-5.3 are judged largely by the labs that build or test them. The exploit count is the one number outsiders can watch directly. It went up by one in a week.

Comments 0

Sources

8
  1. 01Anthropic's super bug-hunting model Mythos is hardcore good at math, as latest vuln under attack showsEN
  2. 02Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesEN
  3. 03OpenAI alerts 100+ orgs that its 'misaligned models' attempted to break in - or worseEN
  4. 04Shield: A 118M model for detecting prompt injections and jailbreaksEN
  5. 05New in llama.cpp: Decision ModelsEN
  6. 06Kolibri Has Landed: A Sovereign Open-Weight ModelEN
  7. 07Changes to Gemini model access and limitsEN
  8. 08Ideogram 4.5: The most precise edit modelEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.