Note:
This vulnerability was originally reported to the Google OSS VRP (Issue ID: 477542544). The Google Security Team requested that I coordinate directly with the Envoy maintainers for triage and remediation. I am submitting this report here to facilitate that process.
Technical Details
I have identified a logic vulnerability in Envoy's HTTP connection manager (FilterManager) that allows for Zombie Stream Filter Execution. This issue creates a "Use-After-Free" (UAF) or state-corruption window where filter callbacks are invoked on an HTTP stream that has already been logically reset and cleaned up.
Mechanism:
The vulnerability resides in source/common/http/filter_manager.cc within the FilterManager::decodeData method.
When an HTTP/2 stream encounters a reset condition (e.g., StreamIdleTimeout, OverloadManager limits, or a local reset triggered by a filter), Envoy calls onResetStream. This method:
- Sets the internal state
state_.saw_downstream_reset_ = true.
- Invokes
onDestroy() on all filters in the chain (allowing them to release resources/pointers).
- Schedules the
ActiveStream object for deferred deletion (cleanup happens later in the event loop).
The Flaw:
The ActiveStream object remains valid in memory during the deferred deletion window. If a DATA frame arrives on this stream immediately after the reset (e.g., in the same packet processing cycle), the HTTP/2 codec invokes ActiveStream::decodeData, which cascades to FilterManager::decodeData.
FilterManager::decodeData fails to check the saw_downstream_reset_ flag. It iterates over the decoder_filters_ list and invokes decodeData() on filters that have already received onDestroy().
Root Cause Code Location:
File: source/common/http/filter_manager.cc
Function: FilterManager::decodeData
void FilterManager::decodeData(...) {
if (stopDecoderFilterChain()) { return; }
// Vulnerability: Missing check for state_.saw_downstream_reset_
// Execution proceeds into the loop even if the stream is logically dead.
auto trailers_added_entry = decoder_filters_.end();
for (; entry != decoder_filters_.end(); entry++) {
// ... calls (*entry)->handle_->decodeData(data) on destroyed filters ...
}
}
Suggested Fix:
Add an explicit state check at the beginning of FilterManager::decodeData.
// Prevent execution on streams that have been reset but not yet destroyed.
if (state_.saw_downstream_reset_) {
return;
}
Impact Analysis
Who can exploit this:
Any remote attacker capable of establishing an HTTP/2 or HTTP/3 connection. No privileges/authentication required.
Impact & Gain:
1. Memory Corruption & Potential Remote Code Execution:
While the immediate symptom is a crash (DoS), the underlying primitive is a Use-After-Free (CWE-416).
- Mechanism: When
onDestroy() is called on filters (e.g., Lua, Wasm, or complex native filters), they release internal structures and invalidate pointers.
- Exploitation: By forcing
decodeData() to execute on these now-freed objects, an attacker triggers undefined behavior. In a heap-groomed environment, an attacker could potentially replace the freed filter object with a malicious payload before the "Zombie" decodeData call occurs. This would allow for vtable hijacking or arbitrary write-what-where primitives, leading to Remote Code Execution (RCE).
- Risk Amplification: This is particularly dangerous for Envoy deployments using memory-unsafe extensions or third-party filters (C++ extensions), where
onDestroy logic is relied upon for safety.
2. Security Control Bypass:
The vulnerability defeats Envoy's "Fail-Closed" security architecture.
- Scenario: If a stream is reset due to a security violation (e.g.,
StreamIdleTimeout, OverloadManager rejection, or WAF triggering), this vulnerability allows the attacker to bypass the termination.
- Result: The attacker can force the processing of "Data" frames on a connection that the security policy explicitly attempted to close, allowing malicious payloads to reach deeper into the filter chain or backend services despite the rejection.
Proof of Concept (Unit Test)
Description:
The attached C++ unit test (zombie_stream_poc_test.cc) deterministically reproduces the vulnerability. It creates a stream, manually triggers a reset (simulating an Overload), and then immediately injects a DATA frame. The test asserts that the filter's decodeData callback is invoked on the reset stream.
#include "test/common/http/conn_manager_impl_test_base.h"
#include "gmock/gmock.h"
#include "gtest/gtest.h"
using testing::_;
using testing::Invoke;
using testing::NiceMock;
using testing::Return;
namespace Envoy {
namespace Http {
/**
* Proof of Concept for "Zombie Stream Filter Execution" (HTTP/2 Reset Re-entrancy)
* * Logic flow:
* 1. Open a stream with HEADERS.
* 2. Force a stream reset (simulating an Overload or Timeout).
* 3. Immediately inject DATA into the stream.
* 4. ASSERT that the filter's decodeData is called despite the stream being reset.
*/
class ZombieStreamPocTest : public HttpConnectionManagerImplTest {
};
TEST_F(ZombieStreamPocTest, ReproducedZombieFilterExecution) {
setup(SetupOpts().setTracing(false));
// 1. Setup a mock filter
std::shared_ptr<MockStreamDecoderFilter> filter(new NiceMock<MockStreamDecoderFilter>());
// Vuln confirmation:
// We expect decodeData to be called on this filter even though the stream is reset.
// In a secure/patched implementation, this EXPECT_CALL should fail (Times(0)).
EXPECT_CALL(*filter, decodeData(_, _))
.Times(1)
.WillOnce(Invoke([&](Buffer::Instance&, bool) -> FilterDataStatus {
ENVOY_LOG_MISC(error, "!!! VULNERABILITY REPRODUCED: decodeData called on a reset stream !!!");
return FilterDataStatus::Continue;
}));
EXPECT_CALL(*filter, decodeHeaders(_, false))
.WillOnce(Return(FilterHeadersStatus::StopIteration));
// Register the filter
EXPECT_CALL(filter_factory_, createFilterChain(_))
.WillOnce(Invoke([&](FilterChainFactoryCallbacks& callbacks) -> bool {
auto factory = createDecoderFilterFactoryCb(filter);
callbacks.setFilterConfigName("vulnerable_filter");
factory(callbacks);
return true;
}));
// 2. Start the stream
EXPECT_CALL(*codec_, dispatch(_))
.WillOnce(Invoke([&](Buffer::Instance&) -> Http::Status {
decoder_ = &conn_manager_->newStream(response_encoder_);
RequestHeaderMapPtr headers{new TestRequestHeaderMapImpl{
{":authority", "host"}, {":path", "/"}, {":method", "POST"}}};
decoder_->decodeHeaders(std::move(headers), false);
return Http::okStatus();
}));
// Dispatch headers
Buffer::OwnedImpl header_buffer("headers");
conn_manager_->onData(header_buffer, false);
// 3. Trigger a Reset on the ActiveStream
// This simulates Envoy terminating the stream due to an external event (Overload, Timeout).
auto* active_stream = dynamic_cast<ConnectionManagerImpl::ActiveStream*>(decoder_);
// This sets state_.saw_downstream_reset_ = true and triggers filter->onDestroy()
active_stream->onResetStream(StreamResetReason::LocalReset, "simulated_overload");
// 4. Attack: Send DATA to the "Zombie" stream
// The ActiveStream object is still alive in the deferred delete list.
Buffer::OwnedImpl malicious_payload("attacker_data");
// This call reaches the filter because FilterManager::decodeData misses the check!
active_stream->decodeData(malicious_payload, false);
}
} // namespace Http
} // namespace Envoy
Note:
This vulnerability was originally reported to the Google OSS VRP (Issue ID: 477542544). The Google Security Team requested that I coordinate directly with the Envoy maintainers for triage and remediation. I am submitting this report here to facilitate that process.
Technical Details
I have identified a logic vulnerability in Envoy's HTTP connection manager (
FilterManager) that allows for Zombie Stream Filter Execution. This issue creates a "Use-After-Free" (UAF) or state-corruption window where filter callbacks are invoked on an HTTP stream that has already been logically reset and cleaned up.Mechanism:
The vulnerability resides in
source/common/http/filter_manager.ccwithin theFilterManager::decodeDatamethod.When an HTTP/2 stream encounters a reset condition (e.g.,
StreamIdleTimeout,OverloadManagerlimits, or a local reset triggered by a filter), Envoy callsonResetStream. This method:state_.saw_downstream_reset_ = true.onDestroy()on all filters in the chain (allowing them to release resources/pointers).ActiveStreamobject for deferred deletion (cleanup happens later in the event loop).The Flaw:
The
ActiveStreamobject remains valid in memory during the deferred deletion window. If aDATAframe arrives on this stream immediately after the reset (e.g., in the same packet processing cycle), the HTTP/2 codec invokesActiveStream::decodeData, which cascades toFilterManager::decodeData.FilterManager::decodeDatafails to check thesaw_downstream_reset_flag. It iterates over thedecoder_filters_list and invokesdecodeData()on filters that have already receivedonDestroy().Root Cause Code Location:
File:
source/common/http/filter_manager.ccFunction:
FilterManager::decodeDataSuggested Fix:
Add an explicit state check at the beginning of
FilterManager::decodeData.Impact Analysis
Who can exploit this:
Any remote attacker capable of establishing an HTTP/2 or HTTP/3 connection. No privileges/authentication required.
Impact & Gain:
1. Memory Corruption & Potential Remote Code Execution:
While the immediate symptom is a crash (DoS), the underlying primitive is a Use-After-Free (CWE-416).
onDestroy()is called on filters (e.g., Lua, Wasm, or complex native filters), they release internal structures and invalidate pointers.decodeData()to execute on these now-freed objects, an attacker triggers undefined behavior. In a heap-groomed environment, an attacker could potentially replace the freed filter object with a malicious payload before the "Zombie"decodeDatacall occurs. This would allow for vtable hijacking or arbitrary write-what-where primitives, leading to Remote Code Execution (RCE).onDestroylogic is relied upon for safety.2. Security Control Bypass:
The vulnerability defeats Envoy's "Fail-Closed" security architecture.
StreamIdleTimeout,OverloadManagerrejection, or WAF triggering), this vulnerability allows the attacker to bypass the termination.Proof of Concept (Unit Test)
Description:
The attached C++ unit test (
zombie_stream_poc_test.cc) deterministically reproduces the vulnerability. It creates a stream, manually triggers a reset (simulating an Overload), and then immediately injects a DATA frame. The test asserts that the filter'sdecodeDatacallback is invoked on the reset stream.