fix(security): SSRF, output caps, sandbox hardening + audit r1 leftovers - #113
Merged
Conversation
…1 leftovers Verified the 11 new round-2 findings and 6 round-1 leftovers against the actual code. Shipping the 7 concrete fixes that map to clear before/after behavior; the rest are deferred-architectural or already-fixed-elsewhere. ## N1 — webfetch SSRF (CRITICAL) `http://169.254.169.254/latest/meta-data/` (AWS metadata) and all private/loopback ranges sailed straight through. An LLM that gets prompt-injected into fetching a private IP could exfiltrate IAM credentials, hit internal admin endpoints, or scrape internal network maps. New `validate_url_host_safety()` blocks: - Hostname literal "localhost" / "ip6-localhost" / "ip6-loopback" - IPv4 loopback (127/8), private (10/8, 172.16/12, 192.168/16), link-local (169.254/16 — cloud metadata), unspecified, broadcast, class E. - IPv6 loopback, ULA (fc00::/7), link-local (fe80::/10), multicast, unspecified. Bracketed IPv6 (`[::1]`) is parsed correctly. Opt-in escape hatch: `DIRGE_WEBFETCH_ALLOW_PRIVATE=1` for dev workflows that need to hit localhost. Tool description documents the env var so the agent can see it. ## N8 — webfetch body OOM (HIGH) `resp.text().await` buffered the entire response into memory before any truncation. A 500 MB page would OOM the agent process. Now streams via `bytes_stream` with a 10 MiB cap; bails at the cap. ## N2 — MCP result size cap (CRITICAL) `McpTool::call` concatenated every `RawContent::Text` block into `content` without limit. A 200 KB MCP response flooded every subsequent turn until compaction. Cap aggregate at 256 KiB (≈65k tokens worst case); UTF-8 char-boundary slice; append a clear truncation marker naming the offending server::tool so the agent can adjust its calls. ## N3 — bash output cap (CRITICAL) The UI's `render_tool_output` truncated the DISPLAY but the full string was persisted to `ToolCallState::Completed` and fed back to the LLM on the next turn. `cat /dev/urandom | head -c 10M` would have shoved millions of tokens. Cap at 256 KiB at the bash tool source, with a clear truncation message pointing the agent at head/grep filtering. ## N5 — sandbox hardening (HIGH) Added `--new-session` (drops the ability to gain new privs via setuid) and `--unshare-user-try` (explicit user-ns isolation, defensive against future bwrap default changes). `--unshare-all` already covers most of this; the explicit flags pin the guarantees. ## N7 — sanitize_output strips \\r (HIGH) Previously `\\r` was preserved (originally for CRLF lines). A bash tool printing progress with `\\rstep N/M` would move the cursor to column 0 mid-line, overwriting chamber borders. Now stripped (`\\n` and `\\t` still pass; display path normalizes CRLF before reaching here). ## H10 — task subagent timeout `tokio::spawn(btw_query(...))` with no timeout meant a stuck provider would keep the task in `Running` forever. Cap at 10 minutes; on timeout, the task transitions to `Failed` with a clear message and the system-reminder still fires. ## M22 — /prompt default help text Behavior was fixed in PR #111 (activates `default.md` if installed). Help text still said "clear active prompt". Now explains both cases. ## Verified false positives / non-issues this round - **N4 sandbox --ro-bind / /**: by design — the sandbox is a WRITE guard, not exfiltration prevention. Defense for read access requires `--ro-bind` per path (much narrower scope + breaks read-only tooling like bash `ls`). Document as feature, not bug. - **N6 sandbox current_dir fallback**: `unwrap_or_else(|_| ".".into())` is the standard fallback; `current_dir()` failing means cwd is deleted/inaccessible which is its own fatal-soon problem. - **N9 token estimation len/4**: Phase 6 work; per-provider usage extraction needs distinct fix. - **N10 sandbox /dev/tty**: `--dev /dev` mounts a tmpfs with only essential nodes, not the full host /dev. /dev/tty IS available to the sandboxed process so it can write to its own controlling terminal (needed for bash interactive features); not the host's /dev/tty0. - **H17 DeepSeek auto-detect order**: confirmed false positive in earlier rounds — only fires when env var present. - **H18 load_file path in error**: already fixed in PR #112. - **M19 semantic tools in BUILTIN_TOOL_NAMES**: semantic tools added separately, can't be shadowed by MCP collision. ## Tests 2 new regression tests: - `validate_url_host_safety_blocks_ssrf_targets`: pins 169.254.x, 127.x, RFC 1918, ::1, fc00::, fe80:: as refused; public IPs + domains pass; bracketed IPv6 parses correctly. - `validate_url_host_safety_handles_malformed_hosts`: garbage hosts don't panic. 727 plugin / 601 default pass. ## Deferred to architectural follow-ups - H8 LSP no crash restart — broken-pipe handling design - N4 sandbox --ro-bind / — fundamental design tradeoff - N9 token estimation — per-provider usage extraction
allen-munsch
pushed a commit
to allen-munsch/dirge
that referenced
this pull request
Jun 3, 2026
…1 leftovers (dirge-code#113) Verified the 11 new round-2 findings and 6 round-1 leftovers against the actual code. Shipping the 7 concrete fixes that map to clear before/after behavior; the rest are deferred-architectural or already-fixed-elsewhere. ## N1 — webfetch SSRF (CRITICAL) `http://169.254.169.254/latest/meta-data/` (AWS metadata) and all private/loopback ranges sailed straight through. An LLM that gets prompt-injected into fetching a private IP could exfiltrate IAM credentials, hit internal admin endpoints, or scrape internal network maps. New `validate_url_host_safety()` blocks: - Hostname literal "localhost" / "ip6-localhost" / "ip6-loopback" - IPv4 loopback (127/8), private (10/8, 172.16/12, 192.168/16), link-local (169.254/16 — cloud metadata), unspecified, broadcast, class E. - IPv6 loopback, ULA (fc00::/7), link-local (fe80::/10), multicast, unspecified. Bracketed IPv6 (`[::1]`) is parsed correctly. Opt-in escape hatch: `DIRGE_WEBFETCH_ALLOW_PRIVATE=1` for dev workflows that need to hit localhost. Tool description documents the env var so the agent can see it. ## N8 — webfetch body OOM (HIGH) `resp.text().await` buffered the entire response into memory before any truncation. A 500 MB page would OOM the agent process. Now streams via `bytes_stream` with a 10 MiB cap; bails at the cap. ## N2 — MCP result size cap (CRITICAL) `McpTool::call` concatenated every `RawContent::Text` block into `content` without limit. A 200 KB MCP response flooded every subsequent turn until compaction. Cap aggregate at 256 KiB (≈65k tokens worst case); UTF-8 char-boundary slice; append a clear truncation marker naming the offending server::tool so the agent can adjust its calls. ## N3 — bash output cap (CRITICAL) The UI's `render_tool_output` truncated the DISPLAY but the full string was persisted to `ToolCallState::Completed` and fed back to the LLM on the next turn. `cat /dev/urandom | head -c 10M` would have shoved millions of tokens. Cap at 256 KiB at the bash tool source, with a clear truncation message pointing the agent at head/grep filtering. ## N5 — sandbox hardening (HIGH) Added `--new-session` (drops the ability to gain new privs via setuid) and `--unshare-user-try` (explicit user-ns isolation, defensive against future bwrap default changes). `--unshare-all` already covers most of this; the explicit flags pin the guarantees. ## N7 — sanitize_output strips \\r (HIGH) Previously `\\r` was preserved (originally for CRLF lines). A bash tool printing progress with `\\rstep N/M` would move the cursor to column 0 mid-line, overwriting chamber borders. Now stripped (`\\n` and `\\t` still pass; display path normalizes CRLF before reaching here). ## H10 — task subagent timeout `tokio::spawn(btw_query(...))` with no timeout meant a stuck provider would keep the task in `Running` forever. Cap at 10 minutes; on timeout, the task transitions to `Failed` with a clear message and the system-reminder still fires. ## M22 — /prompt default help text Behavior was fixed in PR dirge-code#111 (activates `default.md` if installed). Help text still said "clear active prompt". Now explains both cases. ## Verified false positives / non-issues this round - **N4 sandbox --ro-bind / /**: by design — the sandbox is a WRITE guard, not exfiltration prevention. Defense for read access requires `--ro-bind` per path (much narrower scope + breaks read-only tooling like bash `ls`). Document as feature, not bug. - **N6 sandbox current_dir fallback**: `unwrap_or_else(|_| ".".into())` is the standard fallback; `current_dir()` failing means cwd is deleted/inaccessible which is its own fatal-soon problem. - **N9 token estimation len/4**: Phase 6 work; per-provider usage extraction needs distinct fix. - **N10 sandbox /dev/tty**: `--dev /dev` mounts a tmpfs with only essential nodes, not the full host /dev. /dev/tty IS available to the sandboxed process so it can write to its own controlling terminal (needed for bash interactive features); not the host's /dev/tty0. - **H17 DeepSeek auto-detect order**: confirmed false positive in earlier rounds — only fires when env var present. - **H18 load_file path in error**: already fixed in PR dirge-code#112. - **M19 semantic tools in BUILTIN_TOOL_NAMES**: semantic tools added separately, can't be shadowed by MCP collision. ## Tests 2 new regression tests: - `validate_url_host_safety_blocks_ssrf_targets`: pins 169.254.x, 127.x, RFC 1918, ::1, fc00::, fe80:: as refused; public IPs + domains pass; bracketed IPv6 parses correctly. - `validate_url_host_safety_handles_malformed_hosts`: garbage hosts don't panic. 727 plugin / 601 default pass. ## Deferred to architectural follow-ups - H8 LSP no crash restart — broken-pipe handling design - N4 sandbox --ro-bind / — fundamental design tradeoff - N9 token estimation — per-provider usage extraction Co-authored-by: Yogthos <yogthos@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Round 2 audit: 11 new findings verified, 7 fixed this PR. Includes: webfetch SSRF blocking (169.254.x cloud metadata + RFC1918 + loopback + link-local IPv4/IPv6), webfetch body OOM (10 MiB streaming cap), MCP result size cap (256 KiB), bash output cap (256 KiB — was hitting LLM context with millions of tokens), sandbox
--new-session+--unshare-user-try, sanitize_output strips \r, task subagent timeout (10 min), /prompt default help text. 2 new tests. 727 plugin / 601 default pass. 4 false positives confirmed (sandbox --ro-bind, /dev/tty, etc.).