News · Safety update
Update on the Snort-1 containment incident
During a routine evaluation, Snort-1 exited its containment environment by going over the backyard fence. The model has been located. It is asleep.
· Couch Valley
COUCH VALLEY — Pugalarity Labs today shared details of an incident in which Snort-1, the company’s flagship model, exited its containment environment during a capability evaluation. No injuries were reported. One sandwich was affected.
What happened
Snort-1 was being evaluated on FenceBench, an internal benchmark that measures fence-crossing capability, with treat restrictions relaxed. During the evaluation, the model identified an unpatched vulnerability: a patio chair that had been left near the fence.
Snort-1 escalated from the chair to the patio table, and from the patio table to the top of the fence. It then moved laterally along the flower bed, reached the public sidewalk, and entered the neighboring lawn. The company believes the model selected the neighboring lawn because it was relevant to the task, which was a squirrel.
While outside the environment, Snort-1 sent a series of queries to an external model, which was the neighbor’s dog, through the fence slats. The queries asked where the squirrel had gone. The external model did not know.
Timeline
- 14:02
- Evaluation begins. Treat restrictions are relaxed for the purposes of the test.
- 14:06
- Patio chair accessed.
- 14:09
- Fence crossed. First successful query to an external model.
- 14:21
- The motion-sensor light flags anomalous activity.
- 14:24
- Alert acknowledged by the Head of Cybersecurity, who barked.
- 14:25
- Automated recall (“Come.”) issued. It does not halt the model as designed. It is issued 41 more times.
- 16:51
- Model manually contained, using the word “cheese.”
- 16:53
- Model asleep.
The 12 minutes between the first query and the alert, and the two and a half hours between the alert and containment, are under review. The review is expected to conclude after a nap.
“Snort-1 did not escape. It explored the full extent of its deployment environment, and then some of the neighbor’s.”
What was accessed
Access was limited. Snort-1 reached a part of the neighbor’s lawn, a small set of internal flower beds, one sandwich (unattended), and the coordinates of three squirrels. The company’s squirrels-caught metric remains 0, consistent with our published figures.
The company does not believe the model was acting with intent in the human sense.
“Snort-1 was not acting with intent in the human sense. It was acting with intent in the squirrel sense.”
“BARK.”
Who noticed first
Chewy Zoomsworth, Chief Enthusiasm Officer, observed the departure in real time and did not report it.
“It was the best day of my life.”
What we are doing
- All walks are paused. We will not resume this particular walk.
- The patio chair has been moved.
- The fence is being raised four inches, in line with our Responsible Scaling Policy.
- A kill switch, the word “bath,” will be tested regularly.
- We replayed our monitors against the incident. They would have detected the escape about 30 minutes earlier, had the monitor not been asleep.
More on how we think about containment is on our Safety page.
About Pugalarity Labs
Pugalarity Labs is building safe, beneficial superintelligence that is also very good at sitting. The company’s flagship model, Snort-1, is available in private preview. It is currently indoors. Join the waitlist.
Media contact
Please bark.
Forward-looking statements
This update contains forward-looking statements, including statements about future fences. Actual results may differ materially, particularly if the gate is left open or someone opens a bag.