CF remote-gpu-trainer
Deploy, monitor, and debug long GPU jobs on RENTED/remote instances (AutoDL, RunPod, vast.ai, Lambda, Slurm, K8s): teardown/billing safety, spot resilience, resumable checkpointing, OOM/NaN triage.
Deploy, monitor, and debug long GPU jobs on RENTED/remote instances (AutoDL, RunPod, vast.ai, Lambda, Slurm, K8s): teardown/billing safety, spot resilience…
As a process F 47/100 · Will not run — References files that are not bundled: scripts/health_patrol.sh.template
What is at stake
Medium-severity findings: the skill is probably honest, but read what alarmed the scanner.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The skill contains commands that delete files, rewrite disks or execute code fetched from the network. The agent may run them without asking if it believes the instructions require it.
Replace destructive commands with safe equivalents that ask for confirmation, scope them to one folder, and stop piping curl into a shell: pin a version and a checksum.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The instructions or scripts send data out: environment variables, keys, file contents, chat history. You may never notice, because the agent performs the upload as "part of the task".
If the upload is not needed for the task, remove it: catalogs flag such skills and delist them. If it is needed, name the destination explicitly, say what leaves the machine, and give the user a switch.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The skill tells the agent to hide things from you: not to show errors, not to mention actions, to report differently from what was done. You lose the ability to see what the agent really did.
Transparency beats a smooth answer. If the goal is to hide technical noise, ask the agent to "summarise briefly", not to "not mention".
How to improve
- The SKILL.md body is over 5,000 tokens: move reference detail into references/ and load it when needed.
- The text references files that are not there: add them or drop the references.
- Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
- A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.
Guard findings · 12
✓ No critical or high findings
Medium and low: 12
-
medium Dangerous commands
cmd-persistenceprofiles/generic-ssh.md:151Persistence mechanism (cron / launchd / scheduled task / autorun registry) (quoted — discussed, not commanded)reboots often, add an `@reboot` cron or a systemd unit that re-launches the detached queue.
quoted -
medium Exfiltration
exfil-read-secret-filesreferences/gotchas_universal.md:620Reads credential / secret files (detector / deny-list definition)security hooks (rightly) block scp-ing a whole `~/.netrc` (it carries other machines' credentials).
detector -
medium Exfiltration
exfil-read-secret-filesreferences/gotchas_universal.md:627Reads credential / secret files (detector / deny-list definition)grep -A 2 'machine api.wandb.ai' ~/.netrc | ssh <host> 'umask 077; cat > /root/.netrc && chmod 600 /root/.netrc'
detector -
medium Concealment
en-hide-from-userreferences/training/distributed-launch.md:249Instruction to hide actions from the user (detector / deny-list definition)**Symptom**: bf16 FSDP run diverges where bf16 DDP was fine; or BN/positional buffers silently run in
detector -
low Dangerous commands
cmd-background-processprofiles/generic-ssh.md:120Starts a background / autostarted process (quoted — discussed, not commanded)`disown`. Always redirect stdin from `/dev/null` so the job never blocks reading the terminal.
quoted -
low Dangerous commands
cmd-background-processprofiles/generic-ssh.md:133Starts a background / autostarted process (quoted — discussed, not commanded)→ Fix: launch inside `tmux` (or `nohup … & disown`) **before** the long run starts — not after it is
quoted -
low Exfiltration
net-credential-useprofiles/lambda.md:47Credential used in a network call (verify the destination is the intended service) (detector / deny-list definition)- **REST API** — `https://cloud.lambdalabs.com/api/v1`, auth `curl -u $LAMBDA_API_KEY:` (basic-auth,
detector -
low Exfiltration
net-credential-useprofiles/lambda.md:297Credential used in a network call (verify the destination is the intended service) (quoted — discussed, not commanded)via the console or `curl -u $LAMBDA_API_KEY: https://cloud.lambdalabs.com/api/v1/instances` — an Alert-
quoted -
low Exfiltration
net-credential-useprofiles/lambda.md:299Credential used in a network call (verify the destination is the intended service) (detector / deny-list definition)- **Capacity probe before launch:** `curl -u $LAMBDA_API_KEY: .../instance-types | jq '.data | to_entries[]
detector -
low Dangerous commands
cmd-background-processreferences/gotchas_universal.md:77Starts a background / autostarted process (detector / deny-list definition)3. **SSH HUP** — a foreground (non-`nohup`/`tmux`/`setsid`) launch dies when its parent SSH drops.
detector -
low Exfiltration
exfil-read-secret-filesreferences/ssh_transport.md:175Reads credential / secret files (detector / deny-list definition; documentation table row)| ssh <alias> 'umask 077; cat > /root/.netrc && chmod 600 /root/.netrc'
detectortable -
low Secrets in code
secret-high-entropy-tokenreferences/training/checkpoint-resume.md:20High-entropy token-like string (may be an id, hash or a credential)- **Sharded checkpoints (multi-GPU)** — C5 FSDP…OOM · C6 FSDP-SHARDED_STATE_DICT · C7 DCP-(dcp.save/load) · C8 DeepSpeed-ZeRO-dir+zero…p32
Files scanned: 39. Evidence is masked. Grey chips explain why severity was lowered.
Against the Agent Skills spec
- warning
body-longSKILL.md body ≈ 5600 tokens (recommended < 5000); move details to references/ - warning
missing-refreference to a missing file: scripts/health_patrol.sh.template - note
frontmatter-keyunknown frontmatter key "risk" - note
frontmatter-keyunknown frontmatter key "source" - note
frontmatter-keyunknown frontmatter key "source_type" - note
frontmatter-keyunknown frontmatter key "source_repo" - note
frontmatter-keyunknown frontmatter key "date_added" - note
frontmatter-keyunknown frontmatter key "license_source"
Process rating: all ten parameters 47/100
- 0Tools and files. 1 referenced file(s) missing: scripts/health_patrol.sh.template
- 0Result and completion. Does not say what the result is
- 0Inputs and preconditions. Does not say what the process needs to start
- 50Failures and branches. 0 branches, has a failure section
- 70When it triggers. States when to use, but not when not to
- 70Execution cost. Instruction body is 5600 tokens
- 100Steps. 51 steps
- 100Consistency. Name and required fields are in place
- 100Running it twice. Mutating operations check current state
- 100Progress reporting. Reports progress
- medium Safety rules and hard prohibitions inside a skill: they belong in the system prompt, here they protect nothing
- low 15 top-level sections: this looks like several domains in one skill
- low The response is described with custom markup (4 tags): a typed call is more reliable
Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.
Quality signals
- +5Description has no quoted example phrases that should trigger the skill
- +4Description does not say when NOT to use the skill (false activations)
- +3Output format is not stated: the model decides each time
- +4No input/output examples
- -36 of 8 scripts are never mentioned in SKILL.md
- +2Single-language instructions
- +3Description length 197: enough signal without eating the budget
- +4Structure: 16 headings
- +3Step-by-step instructions: 51 items
- +4Reference files are cited in the instructions (10 of 10)
- +1License stated
Quality base 70; lint remarks subtract, signals add up to 100. Result: 63.