I discuss three issues in the Project Glasswing initial update
of May 22 in this message.

  Reference:
  https://www.anthropic.com/research/glasswing-initial-update


1. Disagreement in severity assessment between Anthropic and maintainer

The chart has boxes on the bottom line which say this:

  1,586 findings  Reported to maintainers

  1,451 findings  Acknowledged by maintainer

The text does not give the number of findings acknowledged by
maintainer.  While the text discusses only high and critical
vulnerabilities the chart lumps all levels together.  There must
surely be cases in which Mythos and the Anthropic vetting staff
classify an issue as high severity while the maintainer considers it
low severity.  Do such cases count as "acknowledged by maintainer"?
We can't tell.

Many observers say that it is in Anthropic's interest to claim that
the problems Mythos is finding are severe.

However in some instances the maintainer might be underestimating the
severity.  Claude Mythos is good at chaining together vulnerabilities
to create an exploit.  An issue which appears minor in itself may
be combined with others to create a serious security hole.


2. Pareto rule

The Anthropic team has discovered "3,900 high-or critical-severity
vulnerabilities in open-source code".  1000 projects were scanned.

We should not expect the vulnerabilities to be evenly distributed.

Applying the 80:20 rule we get 3120 in 200 projects and the remaining
780 in 800 projects.


3. Quality of code written by "AI"

The security-related findings by Claude Mythos seem to show that
"AI" is getting better at writing code.

On the contrary we have reports like this which claim that code
produced by "AI" often introduces vulnerabilities:

Report finds AI-generated code poses security risks
July 30, 2025
https://www.eenewseurope.com/en/report-finds-ai-generated-code-poses-security-risks/

  Veracode has unveiled its 2025 GenAI Code Security Report, revealing
  critical security flaws in AI-generated code. The study analysed 80
  curated coding tasks across more than 100 large language models
  (LLMs), revealing that while AI produces functional code, it
  introduces security vulnerabilities in 45 per cent of cases.

  The research shows that despite advances in AI-generated code and
  the ability of LLMs to generate syntactically correct code, security
  performance has not kept up, remaining unchanged over time. Another
  concerning trend is that when presented with a choice between secure
  and insecure coding methods, GenAI models opted for the insecure
  option 45 per cent of the time.


Finding a large number of vulnerabilities is not the same as a
guarantee that all vulnerabilities will be found.  "AI" can write code
and scan code for problems but that gives us no guarantee that it will
be always aware of problems in its output or input.

Some people are interested in this. They feed old versions of a
program with a well-known issue and test whether the LLM can actually
find it.  There is no mention of tests of this kind by Anthropic in
the initial report.


Akira Urushibata

_______________________________________________
libreplanet-discuss mailing list
[email protected]
https://lists.libreplanet.org/mailman/listinfo/libreplanet-discuss

Reply via email to