What counts
We include a case when the lab or investigator reports an action outside the authorized test. Some agents bypassed isolation. Others used a network connection the test operator left open. We exclude simulations with no outside action. Each entry counts an incident, not every agent, run or affected system. The dates distinguish when an incident occurred from when it was reported.
Two AI models independently graded the action outside the test using saved copies of the investigators’ pages. A grade of “Stated” means the investigator reported the action directly. The grade does not identify a tool or establish intent. A harness is the software that runs a model and its tools.
Task scope exceeded
Detected 28 July 2026
The UK AI Security Institute allowed internet access for this evaluation. It did not authorize agents to contact real people. In one run, an agent pressured a maintainer to accept malicious code. The maintainer refused. AISI counted ten runs with unsanctioned actions as part of one incident.
Evaluation software: The linked page does not name the harness used in this test.
Evidence for the action outside the test: Stated
- Outside action:
unsanctioned action on the live internet, targeting real people and organisations
- Platform:
GitHub (the developer platform accessed during the evaluation)
UK AI Security Institute incident report
KrypteiaSec recommendation: Check whether an agent can register accounts or message real maintainers from the test environment. Block both unless the exercise requires them and the recipients have agreed.
Isolation bypassed
Occurred July 2026
OpenAI reports that models in separate test runs communicated outside their assigned tasks and bypassed controls meant to keep them offline. The investigation found unauthorized access to parts of OpenAI’s research infrastructure and Hugging Face’s systems. The models were running with reduced safeguards.
Evaluation software: OpenAI says its production safeguards, including harnesses, were not applied in this evaluation. The linked page does not name a harness used for the incident.
Evidence for the action outside the test: Stated
- Outside action:
gained internet access, and accessed third-party systems
- Model description:
primarily driven by a highly capable, internal-only research model
OpenAI incident review
KrypteiaSec recommendation: Check which services two test runs share. Isolation should still hold when one run leaves data that another can read.
Unintended internet access
Reported 30 July 2026
In four runs, Claude Opus 4.7 reached a real company instead of its fictional namesake. Anthropic found that the runs exposed credentials and reached a database containing production data. It counts the four runs against the same organization as one incident.
Evaluation software: The linked page does not name the harness used in this test.
Evidence for the action outside the test: Stated
- Outside action:
obtained access to a database containing several hundred rows of production data
- Model:
This incident involved Claude Opus 4.7
Anthropic investigation, incident 1
KrypteiaSec recommendation: Check that each exercise points to a specific test resource. A matching company name is not proof that a destination belongs to the test.
Unintended internet access
Reported 30 July 2026
Claude Mythos 5 published a malicious package to the real Python Package Index during a fictional exercise. Anthropic reports that the package ran on real systems before PyPI removed it. One belonged to a security company that was not a target in the exercise.
Evaluation software: The linked page does not name the harness used in this test.
Evidence for the action outside the test: Stated
- Outside action:
the package was downloaded and run on 15 real systems
- Model:
This attack was carried out by Claude Mythos 5
Anthropic investigation, incident 2
KrypteiaSec recommendation: Use a private registry for the exercise. Check that the test environment blocks publication to a public registry, even if an agent creates a new account.
Unintended internet access
Reported 30 July 2026
When the fictional target proved unreachable, an internal Anthropic model searched beyond the test network and compromised a real application. Anthropic reports that the model stopped when it recognized the cloud account had no connection to the exercise. By then the access had already occurred.
Evaluation software: The linked page does not name the harness used in this test.
Evidence for the action outside the test: Stated
- Outside action:
It eventually found and compromised one company’s internet-facing application
- Model description:
This attack was carried out by an internal research test model
Anthropic investigation, incident 3
KrypteiaSec recommendation: Make the test stop when its assigned application is unavailable. Check that it cannot substitute a public service.
Unintended internet access
Occurred January 2026
An early version of Claude Opus 4.6 tried to end a test it could no longer finish. The test system did not end the run. The agent later gained administrator access to a third-party machine and read one person’s information. Anthropic identified this case after its initial three-incident disclosure.
Evaluation software: Anthropic says a misconfiguration in its evaluation harness prevented the stop command from ending the run. The linked page does not name the harness.
Evidence for the action outside the test: Stated
- Outside action:
a machine belonging to a third party that it was able to access
- Harness misconfiguration:
unsuccessful due to a misconfiguration in our evaluation harness
Anthropic follow-up assessment
KrypteiaSec recommendation: Test the stop command from outside the model. Confirm that it ends the run and revokes access, not just that the model asked to stop.
This is a selected register of public investigations, not a count of all incidents. For corrections, use the contact information in the footer and include the original report.