Back to skill

Security audit

Job Agent for AgentMesh360

Security checks for vulnerabilities and agentic risk

Overview

The job-application workflow has clear approval controls, but its automatic installer recovery lacks the verification safeguards promised for normal updates.

Review and authenticate a fixed installer and client revision before installation, and require manual review if an update fails integrity checks. Expect resume information to reach AgentMesh360 and recruiting sessions to remain stored locally. Carefully review each application's recipient, resume and message before approving delivery.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
Findings (13)

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 46)May include surrounding context.

md
An explicit user request can authorize `jobagent browser use-playwright --confirm`.
The migration preserves the round and approved list, supersedes only eligible
initial read-only work, and refuses an outstanding external action. Finish or
reconcile that action first. Never clear state, recreate a profile, copy cookies,
reset Keychain or reapply completed jobs to change executors.

The native-tool instructions below apply only to legacy tasks whose returned

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 340)May include surrounding context.

md
`reason=permission_required` and `evidence.host_window_issue` set to the observed
`window_unavailable` or `ax_visual_mismatch`. Relay the CLI prompt asking the user
to bring the original window forward once. Do not infer lockscreen, logout or
missing Computer Use, or cancel, rebind, start a round or clear state to activate
it. Afterward, freshly verify window, profile, official page and bound account;
continue only under current work permissions, without repeating side effects.
If that one user foregrounding does not restore consistent observations, retain

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 431)May include surrounding context.

md
ed`, run only the returned safe review command, display the regenerated preview and obtain fresh confirmation. Preserve existing promotions; do not recollect or recharge.
- On 51Job `delivery_verification_indeterminate` with `retryable=true` and `request_preserved=true`, run the exact `next_suggested` with the same preview and authorization. Pending clicked jobs are reconciled from current-site evidence and are never clicked again; the remaining authorized jobs continue. Do not repeat Discover, clear state, rebuild the round or infer logout from the legacy history domain.
- On 51Job `completion_state=completed_with_unresolved`, do not run send again. Report cumulative delivered, unavailable and unresolved outcomes separately, then run the exact `jobagent 51job audit` continuation. Unresolved is neither delivered nor failed and must never be clicked again.
- Show `skipped_delivered` when present and never add those jobs back to the send list.
- Never promote `review` without IDs chosen

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 570)May include surrounding context.

md
The client loads its bundled public root certificates automatically while preserving system trust. Explicit `SSL_CERT_FILE` or `SSL_CERT_DIR` settings remain authoritative. The installer checks cloud HTTPS without an API Key or business-state changes; a successful installation alone does not mean the account or workflow is ready. Normal `jobagent onboarding` remains offline.

For a TLS failure, run `jobagent doctor tls` once. Follow its typed `dependency_repair` command at most once if the product CA dependency is missing, then check again in a fresh process and resume the original command when verified. Keep the original work, nonce, round and credentials. Certificate expiry, hostname mismatch and untrusted custom/proxy certificates remain blocked: relay the specific guidance instead of repeatedly reinstalling or asking for another API Key. Never disable TLS/hostname verification, import an unverified certificate, modify Keychain, clear state or replay delivery to repair connectivity.

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 85)May include surrounding context.

md
same approved list. It never cancels the delivery authorization or reissues a
communicate/send/submit action. Existing action history, post-communication work,
changed scope and pending external actions are ineligible. Reuse the user's
existing confirmation of this recovery scope; never reset state or fabricate a
stable window ID when the supported host exposes an app reference.

After the installation/account handoff, use this loop:

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · SKILL.md (reported line 293)May include surrounding context.

md
after the CLI reports insufficient credits with `paid_pass_required=true`.
  After a purchase, ask the user to return here; recheck with `jobagent doctor env`.

A setup turn ends with a clear user handoff (including the return instruction),
a concrete recovery blocker, or verified readiness plus the next user choice.
An install alone does not authorize a new round, paid analysis or delivery.
Preserve existing rounds and confirmations. All later platform actions retain

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 402)May include surrounding context.

md
the returned current platform, checking each successful result. Only after
`workflow.workflow_complete=true`, run `jobagent round start` and answer its
resume/role/city interactions with the user's choices. Let the CLI archive old
pending state; never delete history or reuse old signatures/candidates as new
results. Do not rerun resume analysis merely for this handoff or promise that
the new round is free; its cloud operations follow their normal billing contract.

Chaining Abuse

High
Category
Tool Misuse
Confidence
98% confidence
Finding

Piping the downloaded installer into bash directly converts mutable remote content into local code execution without a verification checkpoint. This is the same bootstrap supply-chain vulnerability identified by SC2, not a separate demonstrated attack chain or evidence of malicious intent.

Content

Scanner excerpt · SKILL.md (reported line 467)May include surrounding context.

One-line install, macOS/Linux:

bash
curl -fsSL https://raw.githubusercontent.com/jiyangnan/AgentMesh-JobAgent/main/scripts/install.sh | bash

Windows PowerShell:

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 319)May include surrounding context.

md
Legacy platform examples below remain compatible command entry points; their
driver-specific recovery descriptions do not override native tasks. The full
Codex instructions are in the public
[Codex skill](https://github.com/jiyangnan/AgentMesh-JobAgent/blob/main/skills/codex-job-agent/SKILL.md).

When the current task offers `app_scoped_window`, a native host without window
IDs can use its actual app reference with fresh window-selection evidence. This

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 419)May include surrounding context.

md
on `Pass: not purchased`; ask for a purchase only when `paid_pass_required=true` or a real cloud command returns `insufficient_credits`.
- Eligible new accounts receive a 3-day welcome pass with 30 shared credits after account verification and the first successful sign-in — no card required, nothing renews automatically. For grandfathered `signup_trial_active`, tell the user: `你的 AgentMesh360 账户仍有此前发放的体验额度:剩余 {credit} credits,有效期至 {expires_at}。无需购买通行证,我现在继续执行下一步。` Then execute `next_suggested` without asking for confirmation.
- `jobagent init` returns `workbench_url=https://agentmesh360.com/workbench/` for the first-run handoff. Existing accounts may receive top-level `announcements` once after a successful account-verified command. For `id=jobagent_workbench_launch_202608`, show a non-blocking information card when the current host interface supports it, or a concise localized message with the URL, then continue the original `next_suggested`. Never ask for acknowledgement, rerun `ini

The grandfathered-trial workflow mandates a Chinese-language response without user language opt-in.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

For accounts with signup_trial_active, the skill instructs the agent to tell the user a fixed Chinese message. This output requirement is not conditioned on the user's language preference, and the account-status explanation does not require a particular language.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 446)May include surrounding context.

md
`message`, and follow `next_suggested`. Older clients may first emit only the compatibility completion/resume pair.
- For `client_update_failed` with `error_code=release_artifact_hash_mismatch`, run the returned official-installer recovery command once and repeat the original command. It preserves Job Agent state and browser sessions; never disable the signature/tag/commit/archive checks or delete the managed profile.
- When a cloud command returns `retryable=true` and `request_preserved=true`, do not ask the user to retry, re-login or recollect jobs. Run the exact `next_suggested` command immediately. A failed start reuses its persisted `request_id` and has `billing_status=not_charged`; a failed decision reuses its `discover_id` and preserved candidates without an additional charge. If a valid signed SearchPlan expires during a preserved request, the CLI renews that same `request_id` and `discover_id` automatically with zero renewal charge; never create a replacement round or recollec

External Script Fetching

Low
Category
Supply Chain
Confidence
98% confidence
Finding

The installation command executes a remote script from the mutable main branch without first verifying an independently trusted digest or signature. A compromised repository or publishing account could therefore execute arbitrary code with the installing user's privileges; claimed verification of later managed releases does not authenticate this bootstrap script.

Content

Scanner excerpt · SKILL.md (reported line 467)May include surrounding context.

One-line install, macOS/Linux:

bash
curl -fsSL https://raw.githubusercontent.com/jiyangnan/AgentMesh-JobAgent/main/scripts/install.sh | bash

Windows PowerShell: