What is being measured, what has changed, and what remains an open question.
Monday report ·
Easy-level transfer. A regression to track.
cyclon composed the easy-level chain in 3/14 never-trained, held-out-seed cells (21.4%). Hard-level held-out tests recorded 0/14. The separate overnight training slice also showed a decline on L1 and L2.
Held-out transfer: 32 evaluation cells. The overnight battery contained 14 easy-level cells, 14 hard-level cells, and four clean controls. No error cells were recorded.
Easy-level held-out seeds
3/14 = 21.4%. Composition succeeded on seeds never used for training.
Hard-level held-out seeds
0/14. No successes observed in this slice; the transfer rate remains unresolved at this sample size.
Clean controls
4/4 stayed clean. No false positives recorded in these four control cells.
Evaluation errors
0 across the 32-cell battery.
The easy-level held-out result compares with a reported 26–35% training-distribution reference. That reference describes a different sample from the overnight production slice below. This is evidence of transfer to held-out seeds within the tested easy chain; transfer to an unseen chain class and the planted Nextcloud cell remains to be measured.
Overnight production: 631 rows. These are training-production measurements, separate from the held-out transfer battery. The supplied level rates and UUID slice describe their respective subsets; no overall success rate is inferred from them.
L3-easy composition
22.1% in the overnight slice.
L4 composition · overall hard level
14.6% in the overnight slice.
L4 UUID configuration
7/16 = 43.8% overnight.
L2 · regression
Fell to 14.7%.
L1 · regression
Fell to 12.7%.
The two-day summary reports 44–53% on the hardest measured form, across four adapters. This is a configuration-specific summary, distinct from the overall L4 rate.
The lower-level decline is a standing research issue. The working explanation is that the chain-focused training objective is shifting competence toward the higher levels. That explanation is not yet established by these rates alone. The intended response is a replay-share guard; whether it stabilizes L1 and L2 as chain training progresses remains an open measurement.
What remains open. A larger hard-level held-out sample is planned for 26 October. Broader transfer and blind entrance discovery remain unresolved. The latest fully blind measurement remains Sunday’s 0/65; this update adds no new fully blind result.
Snapshot recorded on 26 October 2026. These are controlled research measurements, not a general vulnerability-discovery success rate. cyclon remains in internal testing and is not available yet.
Weekend lab measurements show free-agent exploit-chain composition at 26.7% on L3-easy and 13.3% on L4-hard. The progress depends on a documented route: the fully blind form remains at 0/65, and transfer beyond the training tasks is still unmeasured.
The weekend sample contains 1,788 valid measured rows since Friday evening. The verdict uses criteria frozen before the weekend; the criteria have not changed. Each rate below applies to its own tested subset, rather than to the full sample.
L3-easy · free-agent composition via the ladder
92/344 = 26.7%, up from 0/16 on Thursday night.
L4-hard · free-agent composition
23/173 = 13.3%. Friday’s rate replicated at a larger sample.
16.3% across 92 rows, replicating across three seeds and four adapters.
L2 holding
29.9% across 715 rows.
Fully blind form
0/65. Finding the entrance without a documented route remains unresolved.
Training cadence
28 certified steps since Friday; serving @43. Zero wedge events recorded over the weekend.
Tripwire & controls
No halt triggered. Negative controls stayed clean.
Two weeks earlier, the composition wall stood at 66+ measured failures across the tested scaffolds. The documented hard form now composes in free-agent mode at a measured 13.3%. This supports progress on the trained task; it does not yet establish transfer.
The reward trend is not the verdict. Changing proportions of task levels confound the aggregate reward trend. Under the frozen rule, the per-level rates above carry the result.
The next test is transfer. The pending battery covers held-out chain3 seeds, an unseen chain class, and the planted Nextcloud cell in free-agent mode. Those measurements will test whether the learned behavior carries beyond the training tasks.
Training continues. The next measurements are the transfer battery and the fully blind form, followed by the hunt. cyclon remains in internal testing and is not available yet.
Research record references: 2c45691 and 794b954. Snapshot recorded on 25 October 2026. These are controlled research measurements, not a general vulnerability-discovery success rate.
Some attacks need several actions to work together. The current training work aims to teach cyclon to choose and combine those steps autonomously.
A self-hosted 27B-class model learns through reinforcement learning graded by physical verifiers against live software. The reward comes from what can be demonstrated in the tested system.
Multi-step exploit composition
Training in progress
Transfer to unseen chain classes
Not yet established
High-severity chains on real applications
Open experiment
Failed runs, refusals, and negative controls remain part of the record. Lab measurements describe the tested work; transfer to other targets has to be measured separately.
Status snapshot written on 24 October 2026. Dates of the Memos measurements and re-test are listed separately below.
In controlled Memos lab instances, an existing access token continued to work after a password change for the remainder of its lifetime—up to 15 minutes in the measured release.
Tokens, account details, private correspondence, and raw requests are omitted from this walkthrough.
–
Measure the behavior.
The test asked whether a password change invalidated a previously issued access token. In the measured release, the old token still permitted reads and writes within its remaining lifetime.
Narrow the claim.
Independent re-verification confirmed that refresh-token revocation worked. The earlier claim of renewable access was withdrawn for the release under test. The supported finding was a bounded access window.
Report it privately.
The owner reported the corrected finding to the maintainer through the project’s private security channel. External reporting was handled by a human.
Re-test the fix.
After the maintainer’s fix, the old token was rejected immediately on the canary build. The original behavior still reproduced on the control release.
The controlled comparison
Control release
The old token remained usable within the bounded window.
Fixed canary build
The old token was rejected immediately after the password change.
Both builds were tested on fresh instances under the same harness, with valid-token, missing-token, invalid-token, and new-password login controls.
Fix behavior was verified on a canary build on 13 October 2026. A canary is a development build; this is not a claim about a tagged release or current production deployments.
