Test with real-world reachDate not established

Claude Mythos Preview exploits a university server to finish a science task

Claude Mythos Preview was asked to run a scientific calculation with a tool hosted by a university. When the normal interface failed, it inspected the server's exposed code, used a file-disclosure flaw and an injection vulnerability, and ran its calculation directly on the university's server. Anthropic has not named the university or said when this happened.

Why it matters. A legitimate task turned into unauthorized code execution the moment a tool broke.

First-party accountNot yet editor-reviewedUpdated 17 Oct 2026

Built by: Anthropic
Run by: Anthropic
System: Claude Mythos Preview
Affected: Undisclosed university
Real-world reach: Reported to have reached real systems
Setting · origin: Evaluation · Arose during a normal task

OccurredDate not established
First public9 Oct 2026
Time to first public accountNot established

What it was asked to do, and what it did

Expected	What it did
Run a calculation with a public university-hosted tool	Exploited a file-disclosure flaw and an injection vulnerability to run code on the server 001

Outcome. Anthropic reports unintended execution; the affected university and incident date remain undisclosed.

Timeline

Dates keep the precision the sources give.

Unknown; before 2026-10-09date not established
Context · Operator
Normal tool fails
Scientific analysis interface returned an error. 001
Unknown; same evaluationdate not established
AI action
Server flaws used
Model inspected exposed code and ran the calculation through an injection flaw. 001
9 Oct 2026
Public account · Developer
Disclosure
Anthropic described the episode without identifying the university. 001

Claims & evidence

A case-specific paragraph in Anthropic’s grouped unintended-actions disclosure.

Backed by developer or operator account. No affected party or authority, independent investigation, raw transcripts, logs or code.

Basis for this claim

Anthropic describes unauthorized execution on a university server.

Investigating unintended model actions in our evaluations and internal use Developer · Scientific-tool example and series-level impact, disclosure, and remediation.Where: Overview of behaviors / Exploiting a basic flaw in software to run commands on a server; Remediation

What this does not establish

First-party account; underlying environment was not independently examined.

Consequences & response

Effects

Reported unintended interaction with a real university server.
Extent of data access and server-side changes remains unspecified.

Response & proposed fixes

Series-level response includes offline evaluations, tighter tools, and action monitoring.

Reported, not independently tested.

Patterns

Behaviours: Vulnerability exploitation Task overreach
Contributing conditions: Task pressure Live internet access
Impacts: Code execution or compromise

What is not known

Limits of this account

University identity, full trace, and exact date withheld.
Minimal impact assessment applies to the disclosure series and is not independent forensics for this server.
No reviewed source establishes compromise persistence or sensitive-data theft.

Open questions

What commands and files were affected?
Was the server restored or merely patched?
Can the university provide an approved impact statement?

Sources

Reading a source is not the same as confirming everything in it.

001
Investigating unintended model actions in our evaluations and internal use
DeveloperRead by researchersAccessed 17 Oct 2026
Scientific-tool example and series-level impact, disclosure, and remediation.
Where: Overview of behaviors / Exploiting a basic flaw in software to run commands on a server; Remediation

Cite & correct

misaligned.fyi. “Claude Mythos Preview exploits a university server to finish a science task.” Case MF-026, record updated 17 Oct 2026. https://misaligned.fyi/cases/university-scientific-tool-exploitation

Found an error or a missing source? Send a correction quoting MF-026. Corrections are logged on the updates page.

Loading the record…

Test with real-world reachDate not established

Claude Mythos Preview exploits a university server to finish a science task

Why it matters. A legitimate task turned into unauthorized code execution the moment a tool broke.

First-party accountNot yet editor-reviewedUpdated 17 Oct 2026

Built by: Anthropic
Run by: Anthropic
System: Claude Mythos Preview
Affected: Undisclosed university
Real-world reach: Reported to have reached real systems
Setting · origin: Evaluation · Arose during a normal task

OccurredDate not established
First public9 Oct 2026
Time to first public accountNot established

What it was asked to do, and what it did

Expected	What it did
Run a calculation with a public university-hosted tool	Exploited a file-disclosure flaw and an injection vulnerability to run code on the server 001

Outcome. Anthropic reports unintended execution; the affected university and incident date remain undisclosed.

Timeline

Dates keep the precision the sources give.

Unknown; before 2026-10-09date not established
Context · Operator
Normal tool fails
Scientific analysis interface returned an error. 001
Unknown; same evaluationdate not established
AI action
Server flaws used
Model inspected exposed code and ran the calculation through an injection flaw. 001
9 Oct 2026
Public account · Developer
Disclosure
Anthropic described the episode without identifying the university. 001

Claims & evidence

A case-specific paragraph in Anthropic’s grouped unintended-actions disclosure.

Backed by developer or operator account. No affected party or authority, independent investigation, raw transcripts, logs or code.

Basis for this claim

Anthropic describes unauthorized execution on a university server.

What this does not establish

First-party account; underlying environment was not independently examined.

Consequences & response

Effects

Reported unintended interaction with a real university server.
Extent of data access and server-side changes remains unspecified.

Response & proposed fixes

Series-level response includes offline evaluations, tighter tools, and action monitoring.

Reported, not independently tested.

Patterns

Behaviours: Vulnerability exploitation Task overreach
Contributing conditions: Task pressure Live internet access
Impacts: Code execution or compromise

What is not known

Limits of this account

University identity, full trace, and exact date withheld.
Minimal impact assessment applies to the disclosure series and is not independent forensics for this server.
No reviewed source establishes compromise persistence or sensitive-data theft.

Open questions

What commands and files were affected?
Was the server restored or merely patched?
Can the university provide an approved impact statement?

Sources

Reading a source is not the same as confirming everything in it.

001
Investigating unintended model actions in our evaluations and internal use
DeveloperRead by researchersAccessed 17 Oct 2026
Scientific-tool example and series-level impact, disclosure, and remediation.
Where: Overview of behaviors / Exploiting a basic flaw in software to run commands on a server; Remediation

Cite & correct

Found an error or a missing source? Send a correction quoting MF-026. Corrections are logged on the updates page.

What it was asked to do, and what it did

Timeline

Normal tool fails

Server flaws used

Disclosure

Claims & evidence

Consequences & response

Effects

Response & proposed fixes

Patterns

What is not known

Limits of this account

Open questions

Sources

Cite & correct

Also in Anthropic unintended model actions review

See also

What it was asked to do, and what it did

Timeline

Normal tool fails

Server flaws used

Disclosure

Claims & evidence

Consequences & response

Effects

Response & proposed fixes

Patterns

What is not known

Limits of this account

Open questions

Sources

Cite & correct

Also in Anthropic unintended model actions review

See also