Add security-advisories directory.

security-advisories/2014-03-02-0-c++-integer-overflow.md

+Problem
+=======
+
+Integer overflow in pointer validation.
+
+Discovered by
+=============
+
+Ben Laurie <ben@links.org>
+
+Announced
+=========
+
+2014-03-02
+
+Impact
+======
+
+- Remotely segfault a peer by sending it a malicious message.
+- Possible exfiltration of memory, depending on application behavior.
+
+Fixed in
+========
+
+- git commit [f343f0dbd0a2e87f17cd74f14186ed73e3fbdbfa][0]
+- release 0.5.1.1:
+  - Unix: https://capnproto.org/capnproto-c++-0.5.1.1.tar.gz
+  - Windows: https://capnproto.org/capnproto-c++-win32-0.5.1.1.zip
+- release 0.4.1.1:
+  - Unix: https://capnproto.org/capnproto-c++-0.4.1.1.tar.gz
+- release 0.6 (future)
+
+[0]: https://github.com/sandstorm-io/capnproto/commit/f343f0dbd0a2e87f17cd74f14186ed73e3fbdbfa
+
+Details
+=======
+
+*The following text contains speculation about the exploitability of this
+bug. This is provided for informational purposes, but as such speculation is
+often shown to be wrong, you should not rely on the accuracy of this
+section for the safety of your service. Please update your library.*
+
+A specially-crafted pointer could escape bounds checking by triggering an
+integer overflow in the check. This causes the message to appear as if it
+contains an extremely long list (over 2^32 bytes), stretching far beyond the
+memory actually allocated to the message. If the application reads that list,
+it will likely segfault, but if it manages to avoid a segfault (e.g. because
+it has mapped a very large contiguous block of memory following the message,
+or because it only reads some parts of the list and not others), it could end
+up treating arbitrary parts of memory as input. If the application happens to
+pass that data back to the user in some way, this problem could lead to
+exfiltration of secrets.
+
+The pointer is transitively read-only, therefore it is believed that this
+vulnerability on its own CANNOT lead to memory corruption nor code execution.
+
+This vulnerability is NOT a Sandstorm sandbox breakout. A Sandstorm app's
+Cap'n Proto communications pass through a supervisor process which performs a
+deep copy of the structure. As the supervisor has a very small heap, this
+will always lead to a segfault, which has the effect of killing the app, but
+does not affect any other app or the system at large. If somehow the copy
+succeeds, the copied message will no longer contain an invalid pointer and
+so will not harm its eventual destination, and the supervisor itself has no
+secrets to steal. These mitigations are by design.
+
+Preventative measures
+=====================
+
+In order to gain confidence that this is a one-off bug rather than endemic,
+and to help prevent new bugs from being added, we have taken / will take the
+following preventative measures going forward:
+
+1. A fuzz test of each pointer type has been added to the standard unit test
+   suite. This test was confirmed to find the vulnerability in question.
+2. We will additionally add fuzz testing with American Fuzzy Lop to our
+   extended test suite. AFL was used to find the original vulnerability. Our
+   current tests with AFL show only one other (less-critical) vulnerability
+   which will be reported separately ([2014-03-02-2][2]).
+3. In parallel, we will extend our use of template metaprogramming for
+   compile-time unit analysis (kj::Quantity in kj/units.h) to also cover
+   overflow detection (by tracking the maximum size of an integer value across
+   arithmetic expressions and raising an error when it overflows). Preliminary
+   work with this approach successfully detected the vulnerability reported
+   here as well as one other vulnerability ([2014-03-02-1][3]).
+   [See the blog post][4] for more details.
+4. We will continue to require that all tests (including the new fuzz test) run
+   cleanly under Valgrind before each release.
+5. We will commission a professional security review before any 1.0 release.
+   Until that time, we continue to recommend agaisnt using Cap'n Proto to
+   interpret data from potentially-malicious sources.
+
+I am pleased that measures 1, 2, and 3 all detected this bug, suggesting that
+they have a high probability of catching any similar bugs.
+
+[1]: https://github.com/sandstorm-io/capnproto/tree/master/security-advisories/2014-03-02-0-all-cpu-amplification.md
+[2]: https://github.com/sandstorm-io/capnproto/tree/master/security-advisories/2014-03-02-1-c++-integer-underflow.md
+[3]: https://capnproto.org/news/2015-03-02-security-advisory-and-integer-overflow-protection.html

security-advisories/2014-03-02-1-c++-integer-underflow.md

+Problem
+=======
+
+Integer underflow in pointer validation.
+
+Discovered by
+=============
+
+Kenton Varda <kenton@sandstorm.io>
+
+Announced
+=========
+
+2014-03-02
+
+Impact
+======
+
+- Remotely segfault a peer by sending it a malicious message.
+- Possible exfiltration of memory, depending on application behavior.
+- If the application performs a sequence of operations that "probably" no
+  application does (see below), possible memory corruption / code execution.
+
+Fixed in
+========
+
+- git commit [26bcceda72372211063d62aab7e45665faa83633][0]
+- release 0.5.1.1:
+  - Unix: https://capnproto.org/capnproto-c++-0.5.1.1.tar.gz
+  - Windows: https://capnproto.org/capnproto-c++-win32-0.5.1.1.zip
+- release 0.4.1.1:
+  - Unix: https://capnproto.org/capnproto-c++-0.4.1.1.tar.gz
+- release 0.6 (future)
+
+[0]: https://github.com/sandstorm-io/capnproto/commit/26bcceda72372211063d62aab7e45665faa83633
+
+Details
+=======
+
+*The following text contains speculation about the exploitability of this
+bug. This is provided for informational purposes, but as such speculation is
+often shown to be wrong, you should not rely on the accuracy of this
+section for the safety of your service. Please update your library.*
+
+A `Text` pointer, when non-null, must point to a NUL-terminated string, meaning
+it must have a size of at least 1. Under most circumstances, Cap'n Proto will
+reject zero-size text objects. However, if an application performs the
+following sequence, they may hit a code path that was missing a check:
+
+1. Receive a message containing a `Text` value, but do not actually look at
+   that value.
+2. Copy the message into a `MessageBuilder`.
+3. Call the `get()` method for the `Text` value within the `MessageBuilder`,
+   obtaining a `Text::Builder` for the *copy*.
+
+In this case, the `Text::Builder` will appear to point at a string with size
+2^32-1, starting at a location within the Cap'n Proto message.
+
+The `Text::Builder` is writable. If the application decided to overwrite the
+text in-place, it could overwrite arbitrary memory in the next 4GB of virtual
+address space. However, there are several reasons to believe this is unusual:
+
+- Usually, when an application `get()`s a text field, it only intends to
+  read it. Overwriting the text in-place is unusual.
+- Calling `set()` on the field -- the usual way to overwrite text -- will
+  create an all-new text object and harmlessly discard the old, invalid
+  pointer.
+
+Note that even if an application does overwrite the text, it would still be
+hard for the attacker to exploit this for code execution unless the attacker
+also controls the data that the application writes into the field.
+
+This vulnerability is somewhat more likely to allow exfiltration of memory.
+However, this requires the app to additionally echo the text back to the
+attacker. To do this without segfaulting, the app would either need to attempt
+to read only a subset of the text, or would need to have 2^32 contiguous bytes
+of virtual memory mapped into its address space.
+
+A related problem, also fixed in this change, occurs when a `Text` value
+has non-zero size but lacks a NUL terminator. Again, if an application
+performs the series of operations described above, the NUL terminator check
+may be bypassed. If the app then passes the string to an API that assumes
+NUL-terminated strings, the contents of memory after the text blob may be
+interpreted as being part of the string, up to the next zero-valued byte.
+This again could lead to exfiltration of data, this time without the high
+chance of segfault, although only up to the next zero-valued byte, which
+are typically quite common.
+
+Preventative measures
+=====================
+
+This problem was discovered through preventative measures implemented after
+the security problem discussed in the [previous advisory][1]. Specifically, this
+problem was found by using template metaprogramming to implement integer
+bounds analysis in order to effectively "prove" that there are no integer
+overflows in the core pointer validation code (capnp/layout.c++).
+Tentatively, I believe that this analysis exhaustively covers this file.
+The instrumentation has not been merged into master yet as it requires some
+cleanup, but [check the Cap'n Proto blog for an in-depth discussion][2].
+
+This problem is also caught by capnp/fuzz-test.c++, which *has* been
+merged into master but likely doesn't have as broad coverage.
+
+[1]: https://github.com/sandstorm-io/capnproto/tree/master/security-advisories/2014-03-02-0-c++-integer-overflow.md
+[2]: https://capnproto.org/news/2015-03-02-security-advisory-and-integer-overflow-protection.html

security-advisories/2014-03-02-2-all-cpu-amplification.md

+Problem
+=======
+
+CPU usage amplification attack.
+
+Discovered by
+=============
+
+Ben Laurie <ben@links.org>
+
+Announced
+=========
+
+2014-03-02
+
+Impact
+======
+
+- Remotely cause a peer to use excessive CPU time and other resoucres to
+  process a very small message, possibly enabling a DoS attack.
+
+Fixed in
+========
+
+- git commit [b9da55c6b003cca89b25f19d52749461215292d1][0]
+- release 0.5.1.1:
+  - Unix: https://capnproto.org/capnproto-c++-0.5.1.1.tar.gz
+  - Windows: https://capnproto.org/capnproto-c++-win32-0.5.1.1.zip
+- release 0.4.1.1:
+  - Unix: https://capnproto.org/capnproto-c++-0.4.1.1.tar.gz
+- release 0.6 (future)
+
+[0]: https://github.com/sandstorm-io/capnproto/commit/b9da55c6b003cca89b25f19d52749461215292d1
+
+Details
+=======
+
+The Cap'n Proto list pointer format allows encoding a list whose elements are
+claimed each to have a size of zero. Such a list could claim to have up to
+2^29-1 elements while only taking 8 or 16 bytes on the wire. The receiving
+application may expect, say, a list of structs. A zero-size struct is a
+perfectly legal (and, in fact, canonical) encoding for a struct whose fields
+are all set to their default values. Therefore, the application may notice
+nothing wrong and proceed to iterate through and handle each element in the
+list, potentially taking a lot of time and resources to do so.
+
+Note that this kind of vulnerability is very common in other systems. Any
+system which accepts compressed input can allow an attacker to deliver an
+arbitrarily large uncompressed message using very little compressed bandwidth.
+Applications should do their own validation to ensure that lists and blobs
+inside a message have reasonable size. However, Cap'n Proto takes the
+philosophy that any security mistake that is likely to be common in
+naively-written application code is in fact a bug in Cap'n Proto -- we should
+provide defenses so that the application developer doesn't have to.
+
+To fix the problem, this change institutes the policy that, for the purpose of
+the "message traversal limit", a list of zero-sized elements will be counted as
+if each element were instead one word wide. The message traversal limit is an
+existing anti-amplification measure implemented by Cap'n Proto; see:
+
+https://capnproto.org/encoding.html#amplification-attack
+
+Preventative measures
+=====================
+
+This problem was discovered through fuzz testing using American Fuzzy Lop,
+wich identified the problem as a "hang", although in fact the test case just
+took a very long time to complete. We are incorporating testing with AFL into
+our release process going forward.

security-advisories/README.md

+# Security Advisories
+
+This directory contains security advisories issued for Cap'n Proto.
+
+Each advisory explains not just the bug that was fixed, but measures we are taking to avoid the class of bugs in the future.
+
+Note that Cap'n Proto has not yet undergone formal security review and therefore should not yet be trusted for reading possibly-malicious input. Even so, Cap'n Proto intends to be secure and we treat security bugs no less seriously than we would had security review already taken place.
+
+## Reporting Bugs
+
+Please report security bugs to [security@sandstorm.io](mailto:security@sandstorm.io).