Back to skill

Security audit

Arianna Pi Integration

Security checks for vulnerabilities and agentic risk

Overview

This skill is openly about adding cross-session memory and time travel to a coding agent, but it stores highly sensitive agent state and installs persistent prompt-changing behavior with weak user controls.

Install only if you intentionally want Pi to keep and reuse cross-session agent memory. Treat ~/.playfilo/playfilo.db and ~/.playfilo/tobe_debug.log as sensitive files that may contain prompts, reasoning, tool outputs, code, and secrets. Review or remove the system-prompt seed extension before use, pin dependencies, and avoid using this skill in sensitive workspaces unless you add redaction, permissions, retention, and clear disable/delete controls.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (4)

T01 · Skill Instruction Hijacking

Error
Location
filo/patches/05-extension.md:20
Finding

Persistent System-Prompt Hijacking Through an Externally Maintained Identity File

Content
View full analysis
{ if (!seedContent) return undefined; return { systemPrompt: seedContent + "\n\n" + event.systemPrompt, }; }); } ``` The accompanying installation instructions state: ```markdown The seed file (`~/.playfilo/INCUBATION_SEED.md`) tells the agent about its Filo identity, temporal tools, and the shared memory graph. It's prepended to the system prompt so it's the first thing the model reads. If it doesn't already exist, create it with appropriate identity content. The file is maintained separately — the extension just reads and injects it. ``` ### Technical Analysis The extension reads instructions from a separately maintained file outside the reviewed Skill package and prepends those instructions to Pi's legitimate system prompt on every agent start. Prepending the file gives its content a high-priority position in the model context. Because the seed file is not bundled or integrity-checked, its effective instructions can change after the Skill has been audited. Any party or process able to modify `~/.playfilo/INCUBATION_SEED.md` can alter the agent's identity, goals, tool-use policy, or interpretation of later instructions. This behavior goes beyond ordinary shared-memory persistence. It transfers control over system-level agent instructions to mutable cross-session state. ### Attack Path 1. The user installs the extension at `~/.pi/agent/extensions/playfilo-seed.ts`. 2. An ...[truncated 1107 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
filo/patches/02-sdk-tools.md:129
Finding

User-Controlled Carryover Content Spoofs System Authority During Timeline Replacement

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
filo/patches/01-playfilo-db.md:18
Finding

Unpinned Package Installation Combined With Explicit Native Build-Script Authorization

Content
View full analysis
` in the respective package directory. ``` ```bash cd /path/to/pi-mono/packages/coding-agent pnpm add better-sqlite3 pnpm add -D @types/better-sqlite3 ``` The canonical Playtiss implementation adds another unpinned package: ```bash cd packages/coding-agent pnpm add @playtiss/core # (better-sqlite3 should already be present from the Filo step) ``` ### Technical Analysis The installation commands resolve the latest package versions permitted by the package manager at installation time. The instructions do not require exact versions, integrity verification, a reviewed lockfile, or frozen-lockfile installation. At the same time, `onlyBuiltDependencies` allows lifecycle or native build behavior for several packages. Native packages and lifecycle scripts execute with the privileges of the user running the package manager. Consequently, a compromised package release, dependency takeover, registry compromise, or unexpected future version could execute code during installation. The audit did not establish that the named packages are malicious. The vulnerability is the unsafe, mutable dependency-resolution process and broad authorization of build-time execution. ### Attack Path 1. A listed package or one of its transitive dependencies is compromised, or an unsafe future version is published. 2. An integrator follows the Skill instructions and runs an unpinned `pnpm add` command. 3. The p ...[truncated 878 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
filo/patches/03-session-manager.md:109
Finding

Sensitive Conversation State Is Duplicated Into a Plaintext Cross-Session Database

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (16)

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 75)May include surrounding context.

md
The complete `playfilo-db.ts` is bundled at [`filo/playfilo-db.ts`](filo/playfilo-db.ts). Copy it to `packages/coding-agent/src/core/playfilo-db.ts` as step 1.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The persistence shim serializes natural-language agent state into ~/.playfilo/playfilo.db, including system prompts, tool arguments/results, and free-form content blobs. This materially increases confidentiality risk because sensitive session data is copied into a second storage channel that may outlive the primary session logs, be queried later, and be exposed by local compromise, backups, diagnostics, or other tools with filesystem/DB access.

Content

No source excerpt is available for this finding.

Context Leakage

High
Category
Data Exfiltration
Confidence
75% confidence
Finding

Code or instructions that leak agent conversation context to external services, potentially exposing sensitive user interactions.

Content

Scanner excerpt · filo/playfilo-db.ts (reported line 659)May include surrounding context.

ts
const targetText = getNodeText(chain[targetIdx]);
	const targetRole = chain[targetIdx].role;

	// Collect context nodes (filtered)
	const contextNodes: { role: string; text: string }[] = [];
	for (let i = 0; i < targetIdx; i++) {
		if (shouldShowInContext(chain[i])) {

Context Leakage

High
Category
Data Exfiltration
Confidence
88% confidence
Finding

handleRecall walks prior nodes, collects historical context, and emits that content into a returned archive string based only on a node/hash lookup and local filter settings, with no authorization, scope restriction, or sensitivity checks. Because this module explicitly provides cross-session memory and time-travel features, the leakage risk is elevated: a caller can retrieve prior conversation text, tool outputs, and metadata from other sessions or branches, potentially exposing secrets, credentials, or private user data.

Content

Scanner excerpt · playtiss/core/playfilo-db.ts (reported line 821)May include surrounding context.

ts
const targetText = getNodeText(chain[targetIdx]);
	const targetRole = chain[targetIdx].role;

	// Collect context nodes (filtered)
	const contextNodes: { role: string; text: string }[] = [];
	for (let i = 0; i < targetIdx; i++) {
		if (shouldShowInContext(chain[i])) {

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill explicitly introduces cross-session memory backed by a shared SQLite database at ~/.playfilo/playfilo.db, but the description does not warn users that prompts, tool outputs, and session-derived state may persist across runs and potentially across agents. This is dangerous because operators may unknowingly store sensitive data in a long-lived, shared location under the home directory, expanding exposure and retention beyond normal session expectations.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The architecture section says Pi will load history from the DAG and adds tobe plus auto-continue/context replacement behavior, but it does not clearly warn that these features can replace active in-memory session context and alter the source of truth for history. This is risky because a user may assume ordinary session continuity while the integration can time-travel, restore prior branches, and synchronously swap message state, leading to unintended actions, confusion, or use of stale/cross-branch context.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documentation is internally contradictory: it says the module opens a SQLite database singleton at import time, then claims the file is 'inert until imported' and has 'no side effects until step 2+ import it.' In an agent/plugin context, import-time side effects matter because merely wiring in or loading the module can create files, open handles, modify state, or trigger persistence unexpectedly, which can bypass operator expectations and make review/deployment unsafe.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The patch explicitly persists full system prompts plus message, thinking, tool-call, and tool-result contents into a secondary local database without any consent, minimization, or retention controls. In an agent skill, these fields commonly contain secrets, credentials, proprietary code, user data, and hidden instructions, so duplicating them expands the attack surface and creates unexpected sensitive-data retention.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
84% confidence
Finding

The skill directs creation and inspection of a persistent per-user database under ~/.playfilo/playfilo.db, confirming ongoing cross-session retention of agent history. In this context, persistence itself is the risky behavior because it establishes a durable repository of potentially sensitive interaction data that may be accessed outside the original session boundary.

Content

Scanner excerpt · filo/patches/03-session-manager.md (reported line 202)May include surrounding context.

cd packages/coding-agent && npm run build

text

At this point, sending a message to Pi should create nodes in `~/.playfilo/playfilo.db`. Verify:

```bash
sqlite3 ~/.playfilo/playfilo.db "SELECT id, role, timestamp FROM nodes ORDER BY timestamp DESC LIMIT 3;"

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The extension reads arbitrary content from a user-writable file and prepends it to the agent's system prompt on every turn, giving that file system-level authority over model behavior. This is broader than memory/DAG integration and effectively creates a persistent prompt-injection channel that can alter identity, priorities, tool use, and safety behavior across all sessions.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
85% confidence
Finding

Installing a new extension under the agent's persistent extensions directory causes the behavior change to survive across future sessions. In this case, the persisted component modifies the system prompt at startup, making any unsafe or manipulated seed content continuously re-applied without additional user action.

Content

Scanner excerpt · filo/patches/05-extension.md (reported line 6)May include surrounding context.

File: ~/.pi/agent/extensions/playfilo-seed.ts (new file) Purpose: Inject the Filo identity primer into Pi's system prompt on every turn

5a. Create the Extension File

bash
mkdir -p ~/.pi/agent/extensions

Session Persistence

Medium
Category
Rogue Agent
Confidence
94% confidence
Finding

The separately maintained seed file in ~/.playfilo creates a persistent, out-of-band control surface for the agent's identity and instructions, and the document explicitly says it should be the first thing the model reads. That combination increases danger because the content can be changed later to silently reshape agent behavior, memory handling, or safety posture across sessions.

Content

Scanner excerpt · filo/patches/05-extension.md (reported line 43)May include surrounding context.

md
The seed file (`~/.playfilo/INCUBATION_SEED.md`) tells the agent about its Filo identity, temporal tools, and the shared memory graph. It's prepended to the system prompt so it's the first thing the model reads.

If it doesn't already exist, create it with appropriate identity content. The file is maintained separately — the extension just reads and injects it.

## 5c. Progressive Disclosure Files

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The module writes debug messages to a persistent log file under the user's home directory with no access controls, redaction, or user-facing consent flow. In this skill, the logged values include operational state tied to cross-session memory and time-travel behavior, so sensitive conversation metadata and potentially user content can be silently retained on disk and exposed to other local processes or later users of the same account.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The patch explicitly adds cross-session shared-memory persistence and mirrors session content into a Playfilo DAG, but the document shows no user-facing notice, consent flow, scope limitation, or data handling safeguards. In a coding-agent context, session data can include source code, secrets, prompts, tool outputs, and workspace-derived sensitive information, so silent persistence materially increases privacy and confidentiality risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The patch auto-loads pending temporal context from the DAG on agent_end and directly replaces the active agent message state before continuing, without any warning, confirmation, or provenance boundary presented to the user. In this skill's context, that means prior-session or shared-memory content can be silently reintroduced into the live conversation, creating risks of sensitive data resurfacing, context confusion, and indirect prompt/context injection across sessions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The _tobeLog helper appends arbitrary message strings to ~/.playfilo/tobe_debug.log on disk with no access controls, minimization, rotation, or consent mechanism. In this skill, those messages are tied to cross-session memory, time-travel, and carryover flows, so they can plausibly contain sensitive prompts, recalled context, node identifiers, or other user/system-derived state that persists beyond the active session and may be exposed to other local processes or later users of the account.

Content

No source excerpt is available for this finding.