Misc Security fixes#2820
Merged
Merged
Conversation
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
The skills cache fetches from a base URL that may be operator-supplied, and the contents are fed to the model as instructions. Using the plain remote transport allowed any IP — including 127.0.0.1, RFC1918, or 169.254.169.254 (cloud metadata) — to be reached, turning a hostile registry into a metadata-exfiltration vector via prompt injection. Switch to httpclient.NewSafeClient so non-public addresses are refused at dial time (after DNS resolution, defeating DNS rebinding). Tests override the client via TestMain since httptest.NewServer binds to 127.0.0.1.
Replace direct == comparisons with errors.Is so wrapped errors are matched correctly, and replace deprecated os.IsNotExist with errors.Is(err, fs.ErrNotExist). For sql.ErrNoRows the existing == form worked because database/sql.Scan is documented to return it unwrapped (and golangci-lint's errorlint allowlists this pattern), but using errors.Is is consistent with the rest of these files (e.g. bm25_database.go already used errors.Is at line 100) and makes the code resilient if a future caller wraps the error. Note: the project already enables errorlint via .golangci.yml, but errorlint intentionally does not flag these specific patterns (sql.Scan -> sql.ErrNoRows is allowlisted; os.IsNotExist is not in errorlint's scope). No linter change is therefore needed.
…ndler The callback server published the OAuth result via three size-1 channels with blocking sends. Any additional hit on /callback — a stale browser tab, the user clicking back then re-submitting, or any local process on the loopback interface poking the port — would wedge the HTTP handler goroutine forever (until Shutdown), leaking goroutines and TCP connections per stray request, and giving a local attacker a trivial way to pin those resources. Collapse codeCh/stateCh/errCh into a single buffered resultCh and use non-blocking sends via a deliver helper. The first callback wins; later ones are dropped on the floor. WaitForCallback no longer has the two-channel race where a context cancellation between codeCh and stateCh reads would lose the code. Add TestCallbackServer_DuplicateCallbacksDoNotBlock which fires five back-to-back callbacks and verifies (a) every HTTP response completes within the timeout and (b) the first one is still delivered to the waiter.
The skills cache previously persisted every response to disk, even when
the upstream marked the body as Cache-Control: no-store. Since the
cached content is fed to the model as instructions, this was both a
privacy hazard (sensitive content lingering under ~/.cagent) and a
violation of RFC 9111 \u00a75.2.2.5 ("the cache MUST NOT store any part
of either the immediate request or response").
no-cache was also conflated with max-age=0 \u2014 it actually means
"may be stored, but must be revalidated before reuse" (RFC 9111
\u00a75.2.2.4). For the skills cache, which has no conditional-GET
support yet, the safe approximation is to store but report a miss on
the next Get, forcing a fresh fetch.
- Replace parseCacheExpiry with parseCacheControl returning a structured
cacheDirective.
- FetchAndStore now skips the disk write entirely on no-store and
returns the body in-memory.
- Get treats fs.ErrNotExist as a clean miss, but logs other errors
(corrupt JSON, EACCES, \u2026) at debug instead of silently masking them
as a benign refetch trigger.
- Add tests for no-store, no-cache, and the directive precedence.
Self-review caught a regression in the previous commit. Skipping the
disk write entirely on Cache-Control: no-store made remote skills with
that header unreadable, because the consumer at pkg/tools/builtin/skills
reads skill.FilePath directly rather than going through diskCache.Get.
After prefetchFiles returned (silently happy), readFileContent(...)
hit ENOENT and the skill was effectively broken.
Restore the disk write but keep no-store distinct from a normal
response: the entry's metadata expires immediately, so the next Load()
cycle refetches rather than reusing the stored copy. This is no longer
strict RFC 9111 \u00a75.2.2.5 ("the cache MUST NOT store any part"),
which is documented in the function comment along with the future
direction (in-memory cache shared with the reader).
Update TestDiskCache_NoStoreStoresButExpiresImmediately accordingly:
the file must now exist on disk, but Get must still report a miss so
prefetch refetches on the next Load.
…isleading 200 Addresses review feedback on the previous commit (drop stray OAuth callbacks instead of blocking the handler). Even though duplicate callbacks no longer wedge the handler goroutine, they were still receiving an 'Authorization Successful!' page even when their result was silently discarded by deliver(). This was misleading for users who refresh a stale browser tab, and gave a local attacker a way to probe whether a flow is in progress (200 vs 400 on a known nonce). deliver() now returns whether it actually delivered the result. The callback handler responds with HTTP 409 Conflict and an 'Authorization Already Processed' page when deliver() returns false, except for the state-mismatch path where the response stays unchanged so it doesn't leak whether a flow is currently active. TestCallbackServer_DuplicateCallbacksDoNotBlock now also verifies that the second through fifth callbacks return HTTP 409 rather than HTTP 200.
dgageot
force-pushed
the
board/a06d029e834182e5
branch
from
c68ce71 to
4f57a88
Compare
aheritier
approved these changes
May 20, 2026
aheritier
added
area/security
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
3 participants
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.
No description provided.