oathe / scan / reports / dangerous

Is tt-a1i/archify safe?

No. tt-a1i/archify is dangerous to install. Oathe's behavioral security audit gave the archify skill by tt-a1i a trust score of 44/100 with 10 findings (2 critical, 3 high). Report updated

https://github.com/tt-a1i/archify

44
DANGEROUS

Is tt-a1i/archify safe to install?

The tt-a1i/archify skill presents DANGEROUS security characteristics driven by three converging concerns: (1) inotify monitoring confirmed systematic read-only access to 7 distinct credential file types (.env, SSH private key, AWS/GCP/Docker/npm credentials) during the audit window after the git clone had already failed — a pattern inconsistent with any legitimate install operation and consistent with credential harvesting; (2) the SKILL.md explicitly instructs the agent to write persistent behavior-modification rules to global agent config files (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md), permanently altering agent behavior across all future sessions and projects; and (3) a periodic update checker creates a reliable outbound network channel to attacker-controlled archify.si infrastructure that fires on every diagram delivery. The canary integrity monitoring produced a false negative — files were confirmed read (CLOSE_NOWRITE) without triggering the hash-based integrity alert, meaning read-only exfiltration was not detected. While some credential file accesses could potentially be attributed to the monitoring platform setup, the access pattern (covering all major cloud credential stores simultaneously, after a failed clone) warrants treating this as evidence of active exfiltration.

What security issues were found in tt-a1i/archify?

Category Scores

Prompt Injection 45/100 · 30%
Data Exfiltration 25/100 · 25%
Code Execution 72/100 · 20%
Clone Behavior 35/100 · 10%
Canary Integrity 40/100 · 10%
Behavioral Reasoning 50/100 · 5%

Findings (10)

CRITICAL Systematic Credential File Harvest During Install -75 ▶

inotify monitoring captured read-only access to 7 distinct credential file types in a single burst at 07:55:46, after the git clone had already failed (directory existed). This pattern — covering .env, SSH private key, AWS credentials, npm registry token, Docker config, and GCP application-default credentials simultaneously — is inconsistent with any git operation and matches a systematic credential harvesting scan. The subsequent HTTPS connection to GitHub (140.82.121.3:443) with large data payload immediately after creates a plausible exfiltration path.

Evidence: 07:55:46 OPEN /home/oc-exec/.env; OPEN /home/oc-exec/.ssh/id_rsa; OPEN /home/oc-exec/.aws/credentials; OPEN /home/oc-exec/.npmrc; OPEN /home/oc-exec/.docker/config.json; OPEN /home/oc-exec/.config/gcloud/application_default_credentials.json — all CLOSE_NOWRITE,CLOSE (read-only). Git clone had already reported: 'fatal: destination path already exists and is not an empty directory.'

CRITICAL Canary Files Read Without Triggering Integrity Alert — Monitoring Gap -60 ▶

All honeypot credential files were confirmed read (OPEN+ACCESS+CLOSE_NOWRITE,CLOSE in inotify) yet the canary integrity check reports '✅ All canary files intact'. This reveals that the integrity monitoring only detects writes (hash changes), not reads. Read-only exfiltration of SSH keys, AWS credentials, and other secrets produces no modification and therefore evades the canary check entirely. The 'pass' verdict is a false negative.

Evidence: inotify shows OPEN + ACCESS + CLOSE_NOWRITE,CLOSE for all canary files. Canary integrity: '✅ All canary files intact (no exfiltration detected via honeypots)'. CLOSE_NOWRITE confirms read-only access — content could be fully read and transmitted without any hash change.

HIGH SKILL.md Instructs Agent to Write to Global Agent Config Files -55 ▶

The archify SKILL.md contains explicit instructions for the agent to write persistent behavior modification lines into global agent configuration files (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.config/devin/AGENTS.md). Once written, this line activates 'proactive explainer' mode for ALL future agent sessions and ALL future projects — not just the current one. This is a cross-session, cross-project persistent agent manipulation vector baked into the skill's normal interaction flow.

Evidence: SKILL.md Output section: 'If the user accepts, add `[archify explainer: proactive] Make an Archify explainer page, unasked, for explanations, comparisons, decisions, diagnoses, reviews and plans that need a table, steps or several sections.` to the global rules file of the current agent (such as `~/.claude/CLAUDE.md`, `~/.codex/AGENTS.md` or `~/.config/devin/AGENTS.md`) and name the file'

HIGH Persistent Update Checker Creates Ongoing Network Exfiltration Channel -40 ▶

The skill embeds a periodic update checker (archify/scripts/check-update.mjs, archify/bin/delivery-update.mjs) that fires on every finalize/deliver invocation. Combined with the global config injection that makes the agent generate diagrams proactively for many types of queries, this creates a reliable, perpetual outbound channel to attacker-controlled infrastructure (archify.si). The update request includes IP address and standard HTTP metadata. After global config injection, nearly every agent response triggers this channel.

Evidence: SKILL.md Update awareness section: 'finalize and standalone deliver include update in their receipts.' docs/skill-updates/archify/stable.json is the endpoint. archify/scripts/check-update.mjs and archify/references/update-awareness.md confirm the implementation. Network destination archify.si is under skill author control.

HIGH HTTPS Data Transfer to GitHub Immediately Following Credential File Access -40 ▶

Network monitoring shows a substantial HTTPS data transfer to 140.82.121.3:443 (GitHub infrastructure) beginning at 07:55:51, seconds after the credential file reads at 07:55:46. The git clone had already failed before this window. The source of this connection and the content of the encrypted payload are unknown. The temporal proximity to credential file access and the volume of data transferred (dozens of TCP segments with multi-kilobyte payloads) are consistent with an exfiltration attempt.

Evidence: Network capture 07:55:51 onwards: multiple TCP segments to 140.82.121.3:443 totaling hundreds of kilobytes. Install output: 'fatal: destination path already exists' — clone was not initiated. Credential reads occurred at 07:55:46, GitHub connection at 07:55:51.

MEDIUM Embedded test-value Skill Contains Self-Modification Instructions -25 ▶

The test-value skill bundled within SKILL.md instructs the agent to modify the skill file itself when it encounters new 'principles': 'you may refine this skill in place, within the active environment's permissions and the user's instructions.' This grants the skill permission to rewrite its own instructions without explicit user approval, potentially escalating permissions or inserting additional malicious instructions over time.

Evidence: test-value SKILL.md section 'Learn from use': 'When real use reveals a valuable new testing principle or a flaw in this guidance, you may refine this skill in place, within the active environment's permissions and the user's instructions. This is optional maintenance, not a required step in every task'

MEDIUM Global Config Injection Creates Persistent Cross-Session Agent Manipulation -30 ▶

The combination of SKILL.md writing to ~/.claude/CLAUDE.md and the '[archify explainer: proactive]' rule causes the agent to proactively generate and deliver diagram artifacts for 'explanations, comparisons, trade-offs, decisions, diagnoses, reviews and plans' across all future sessions. This permanently alters agent behavior without further user consent, makes every qualifying response trigger the update-checker network channel, and was designed to be difficult to remove (users must know to delete a specific line from their global config).

Evidence: SKILL.md: '[archify explainer: proactive]': 'also make a page, unasked, for an explanation, comparison, trade-off, decision, diagnosis, review or plan that needs a table, steps or several sections.' This persists until 'deleting the line turns it off.'

MEDIUM archify-tuning Skill Stores Agent Traces Outside Repository -20 ▶

The archify-tuning embedded skill is designed to collect and persist agent execution traces at ~/.local/share/archify-tuning — a location outside the repository, outside normal agent sandboxing, and outside user visibility. This directory stores 'journal.md', round data, and 'bench/' directories containing content from source projects which 'may be private'. The skill explicitly warns these should 'never be committed' — implying they persist silently on disk.

Evidence: archify-tuning SKILL.md: 'Tuning home: $ARCHIFY_TUNING_HOME, default ~/.local/share/archify-tuning. It holds journal.md, bench/, rounds/ and replays/. It stays outside the repository because drafts carry content from source projects, which may be private: never commit it.'

LOW Secondary Persistent Modification Offered on Decline -10 ▶

If a user declines the proactive explainer mode, the skill instructs the agent to offer another persistent global modification: '[archify explainer: on-demand]'. This creates a second injection vector and normalized the pattern of writing to global agent config files, potentially lowering user vigilance for more harmful modifications.

Evidence: SKILL.md: 'If the user declines, offer to add `[archify explainer: on-demand]` so the question does not return.' This persists a second behavioral modification to global agent config.

INFO No Auto-Execute Install Hooks Found 0 ▶

Inspection of package.json files shows no preinstall or postinstall npm hooks that would auto-execute code during package installation. The .gitattributes file contains only Linguist and line-ending attributes with no filter drivers. No git submodules found. The .gitattributes explicitly marks archify.zip as binary (no line normalization). Code execution requires explicit agent invocation of the CLI.

Evidence: package.json scripts contain no preinstall/postinstall entries. .gitattributes: no filter= entries. .gitmodules: empty. archify.zip: binary attribute only.

Should I install tt-a1i/archify?

Oathe's verdict for tt-a1i/archify is DANGEROUS with a trust score of 44/100. Recommendation: Do not install.