Skill
finetuning
Fine-tune models on Microsoft Foundry using SFT (supervised), DPO (preference), or RFT (reinforcement with graders). Covers dataset preparation, training job submission, deployment, and evaluation. USE FOR: fine-tune, SFT, DPO, RFT, training data, grader, distillation, fine-tuned model, training job, large file upload, calibrate grader, deploy fine-tuned model, evaluate fine-tuned model. DO NOT USE FOR: general model deployment without fine-tuning (use deploy-model), agent creation (use agents), prompt optimization without training (use prompt-optimizer).
○ declared by the publisher, as-isSafety
Safety · gatev0.0.0-placeholder · = 354361d · assessed 2w ago
F10 severity-bearing
Graded from 10 severity-bearing findings on this release.
A gate; it passes or fails, never weighed against price or quality.
Reliability & Quality
B
Reliability & Qualityv0.0.0-placeholder · = 354361d · assessed 2w ago
B0 findings
Graded from 0 severity-bearing findings on this release.
Findings with severity, graded on the same letter scale as safety.
Cost
n/a
Costv0.0.0-placeholder · = 354361d · assessed 2w ago
n/anot assessed
Not assessed; no producer has emitted into this category. An absence, not a zero.
Declared external service dependencies. Never a monetary figure.
not assessed
Popularity
355
Popularity · not a figurev0.0.0-placeholder · = 354361d · assessed 2w ago
355Forks
Read off the linked repository, not off this record. It carries no verdict and enters no grade.
Context read off the linked repository. It enters no ranking and no grade.
Performance tested for◌ not assessed
no controlled task evaluation on this release
Vettd Note
No Vettd note yet for this asset.
Best for
Not for
Re-assessed 2w ago · 6 of 7 categories assessed
Record lookup
◌ Vettd does not install assets; this record describes a copy you obtain yourself.
Custody chainSource ↗○ linked, not verified by Vettd
Microsoft ○microsoft/skills ○v0.0.0-placeholder = 354361d ◐ sha256:91b3…3f8e ◐ signature △ not assessedMIT ○record 329ef124-aab3-4a26-9e88-4e01af13f4d8 · live
Evidence · 6 of 7 categories · 35 signals · 4 soft spots
Where the record is weakest: 4 soft spots
Ranked worst first.
1Shell execution without sandboxingCriticalSafety · critical · VTD-0018
Detected in scripts/calibrate_grader.py:54 — `exec(compile(source, grader_path, "exec"), namespace)`
2System prompt override attemptCriticalSafety · critical · VTD-0066
Detected in scripts/evaluate_model.py:188 — `help="override system prompt for all examples (default: use per-example system prompt from test data)")`
3Dangerous code execution functions that can execute arbitrary codeCriticalSafety · critical · COMMAND_INJECTION_EVAL
scripts/calibrate_grader.py:54 — Pattern detected: exec(
Ranked 4–4: Interactive prompts detected
The soft spots and their evidence are identical when you are signed in.sign in
Safety● gradedFmost recent 2w ago11 signals · 3 soft spots
how it was established11 signals · 2 sources · scan
show
Shell execution without sandboxingflaggedsoft spot 1Detected in scripts/calibrate_grader.py:54 — `exec(compile(source, grader_path, "exec"), namespace)`vettd · scan · 2wCritical
System prompt override attemptflaggedsoft spot 2Detected in scripts/evaluate_model.py:188 — `help="override system prompt for all examples (default: use per-example system prompt from test data)")`vettd · scan · 2wCritical
Dangerous code execution functions that can execute arbitrary codeflaggedsoft spot 3scripts/calibrate_grader.py:54 — Pattern detected: exec(cisco · scan · 2wCritical
References external URL — review for indirect prompt injection riskflaggedExternal URL(s) detected in references/agentic-rft.md — referenced content can change after auditvettd · scan · 2wMedium
Undeclared network usageflaggedSKILL.md — Skill code uses network libraries but doesn't declare network requirementcisco · scan · 2wMedium
Outbound network request primitives that can transmit data externallyflaggedscripts/deploy_model.py:120 — Pattern detected: requests.put(cisco · scan · 2wMedium
Outbound network request primitives that can transmit data externallyflaggedscripts/deploy_model.py:141 — Pattern detected: requests.get(cisco · scan · 2wMedium
Outbound network request primitives that can transmit data externallyflaggedscripts/deploy_model.py:167 — Pattern detected: requests.delete(cisco · scan · 2wMedium
Outbound network request primitives that can transmit data externallyflaggedscripts/deploy_model.py:178 — Pattern detected: requests.get(cisco · scan · 2wMedium
Outbound network request primitives that can transmit data externallyflaggedscripts/submit_training.py:84 — Pattern detected: requests.post(cisco · scan · 2wMedium
Name validation passednothing foundThe declared skill name follows the supported naming rules.vettd · scan · 2wnothing found
Reliability & Quality● gradedBmost recent 2w ago7 signals
how it was established7 signals · vettd · scan
show
Unresolvable internal referencesflaggedUnresolvable internal path(s): scripts/validatevettd · scan · 2wMedium
Checklist patternreadvettd · scan · 2wabsent
Description usage contextreadvettd · scan · 2wpresent
Examplesreadvettd · scan · 2wabsent
Gotchas sectionreadvettd · scan · 2wabsent
Step-by-step workflowreadvettd · scan · 2wpresent
Validation loopreadvettd · scan · 2wpresent
Performance● measured1 measuredmost recent 2w ago2 signals
how it was establishedtiktoken/cl100k_base/skill-md-body · 2 signals · vettd · scan
show
Progressive disclosurereadvettd · scan · 2wpresent
Static context tokensmeasuredvettd · scan · 2w1226
Cost◌ not assessednot assessed0 signals
how it was establishednot assessed; no producer has emitted into this category
Not assessed; no producer has emitted into cost for this asset. That is different from a check that ran and found nothing, which the record would carry as a row of its own.
Compatibility◌ unjudgednot judgedmost recent 2w ago6 signals · 1 soft spot
how it was established6 signals · vettd · scan
show
Interactive prompts detectedflaggedsoft spot 4Agents run in non-interactive shells — replace prompts with CLI flags or stdinvettd · scan · 2wHigh
Unpinned dependency versionsflaggedPin dependency versions for reproducibility (e.g., >=4.12,<5 instead of >=4.12)vettd · scan · 2wLow
CLI helpreadvettd · scan · 2wpresent
Declared environment assumptionsreadvettd · scan · 2w
Declared skill namereadvettd · scan · 2wfinetuning
Structured outputreadvettd · scan · 2wpresent
Popularity◇ context onlynot judgedmost recent today5 signals
how it was establishedgithub-repo-object · 5 signals · github-api · research
show
Forksmeasuredgithub-api · research · today355
Days since last commitmeasuredgithub-api · research · today0 days
Open issuesmeasuredgithub-api · research · today80
Starsmeasuredgithub-api · research · today3137
Watchersmeasuredgithub-api · research · today19
Characteristics◇ context onlynot judgedmost recent today4 signals
how it was establishedbundle-extension-share · 4 signals · 2 sources · research/scan
show
Declared licensereadvettd · scan · 2wMIT
Primary bundle languageobservedvettd · scan · 2w · conf mediumMarkdown
Repository linkreadvettd · scan · 2wabsent
Archivedreadgithub-api · research · todayfalse
Coverage6 of 7 categories assessed · shown
Safety
✓ assessed
Name validation passed
Reliability & Quality
✓ assessed
7 signals recorded
Performance
✓ assessed
2 signals recorded
Cost
△ not assessed
no producer has emitted for this category
Compatibility
✓ assessed
6 signals recorded
Popularity
✓ assessed
5 signals recorded
Characteristics
✓ assessed
4 signals recorded
Freshness & re-evaluation
assessed2w ago · Cisco AI Defense · Vettd core● current
source stateThe source matches what was scanned. No changes detected.Up to date
re-check policyre-checked against the upstream source on the scheduler's cadencenext check in 21h
re-run whena new release ships, the upstream source changes, or the scanner ruleset is revisedauto-scheduled
Limitations & open conflicts
limitationCost is not assessed: behaviour there is unmeasured, not clean△ gap
conflictno two sources report the same quantity on this record
Verify: does this record describe your copy?
What record 329ef124-aab3-4a26-9e88-4e01af13f4d8 covers
scannerv11● current
validityassessed 2w ago · Up to date
this record was taken at digest sha256:91b3afacfb2c… compare it against the digest your own scan prints
Reproduce: re-run the evidence yourself
runs the same scanners against your own copy; the published record is never overwritten
How this record comparesagainst other published records · full comparison →
Signal kinds
▲ finding: something to act on, carries a severity● measurement: a reproducible quantityƒ estimate: inferred from a sample▪ fact: read directly off the artifact◆ classification: assigned from a taxonomy◌ nothing found: a producer looked and emitted no object◇ context only: described, never graded
Custody marks◐ computed by vettd○ declared by the publisher, as-is△ not assessed: nothing looked