Wasmtime is an open source runtime for WebAssembly & WASI. In Wasmtime from version 0.26.0 and before version 0.30.0 is affected by a memory unsoundness vulnerability. There was an invalid free and out-of-bounds read and write bug when running Wasm that uses `externref`s in Wasmtime. To trigger this bug, Wasmtime needs to be running Wasm that uses `externref`s, the host creates non-null `externrefs`, Wasmtime performs a garbage collection (GC), and there has to be a Wasm frame on the stack that is at a GC safepoint where there are no live references at this safepoint, and there is a safepoint with live references earlier in this frame's function. Under this scenario, Wasmtime would incorrectly use the GC stack map for the safepoint from earlier in the function instead of the empty safepoint. This would result in Wasmtime treating arbitrary stack slots as `externref`s that needed to be rooted for GC. At the *next* GC, it would be determined that nothing was referencing these bogus `externref`s (because nothing could ever reference them, because they are not really `externref`s) and then Wasmtime would deallocate them and run `<ExternRef as Drop>::drop` on them. This results in a free of memory that is not necessarily on the heap (and shouldn't be freed at this moment even if it was), as well as potential out-of-bounds reads and writes. Even though support for `externref`s (via the reference types proposal) is enabled by default, unless you are creating non-null `externref`s in your host code or explicitly triggering GCs, you cannot be affected by this bug. We have reason to believe that the effective impact of this bug is relatively small because usage of `externref` is currently quite rare. This bug has been patched and users should upgrade to Wasmtime version 0.30.0. If you cannot upgrade Wasmtime at this time, you can avoid this bug by disabling the reference types proposal by passing `false` to `wasmtime::Config::wasm_reference_types`.

An issue in bytecodealliance wasm-micro-runtime before v.b3f728c and fixed in commit 06df58f allows a remote attacker to escalate privileges via a crafted file to the check_was_abi_compatibility function.

Wasmtime is a runtime for WebAssembly. Prior to 24.0.7, 36.0.7, 42.0.2, and 43.0.1, Wasmtime contains a vulnerability where when transcoding a UTF-16 string to the latin1+utf16 component-model encoding it would incorrectly validate the byte length of the input string when performing a bounds check. Specifically the number of code units were checked instead of the byte length, which is twice the size of the code units. This vulnerability can cause the host to read beyond the end of a WebAssembly's linear memory in an attempt to transcode nonexistent bytes. In Wasmtime's default configuration this will read unmapped memory on a guard page, terminating the process with a segfault. Wasmtime can be configured, however, without guard pages which would mean that host memory beyond the end of linear memory may be read and interpreted as UTF-16. A host segfault is a denial-of-service vulnerability in Wasmtime, and possibly being able to read beyond the end of linear memory is additionally a vulnerability. Note that reading beyond the end of linear memory requires nonstandard configuration of Wasmtime, specifically with guard pages disabled. This vulnerability is fixed in 24.0.7, 36.0.7, 42.0.2, and 43.0.1.

Wasmtime is a runtime for WebAssembly. From 25.0.0 to before 36.0.7, 42.0.2, and 43.0.1, Wasmtime's Winch compiler contains a bug where a 64-bit table, part of the memory64 proposal of WebAssembly, incorrectly translated the table.size instruction. This bug could lead to disclosing data on the host's stack to WebAssembly guests. The host's stack can possibly contain sensitive data related to other host-originating operations which is not intended to be disclosed to guests. This bug specifically arose from a mistake where the return value of table.size was statically typed as a 32-bit integer, as opposed to consulting the table's index type to see how large the returned register could be. When combined with details about Wnich's ABI, such as multi-value returns, this can be combined to read stack data from the host, within a guest. This vulnerability is fixed in 36.0.7, 42.0.2, and 43.0.1.

Wasmtime is a runtime for WebAssembly. From 32.0.0 to before 36.0.7, 42.0.2, and 43.0.1, Wasmtime's Cranelift compilation backend contains a bug on aarch64 when performing a certain shape of heap accesses which means that the wrong address is accessed. When combined with explicit bounds checks a guest WebAssembly module this can create a situation where there are two diverging computations for the same address: one for the address to bounds-check and one for the address to load. This difference in address being operated on means that a guest module can pass a bounds check but then load a different address. Combined together this enables an arbitrary read/write primitive for guest WebAssembly when accesssing host memory. This is a sandbox escape as guests are able to read/write arbitrary host memory. This vulnerability has a few ingredients, all of which must be met, for this situation to occur and bypass the sandbox restrictions. This miscompiled shape of load only occurs on 64-bit WebAssembly linear memories, or when Config::wasm_memory64 is enabled. 32-bit WebAssembly is not affected. Spectre mitigations or signals-based-traps must be disabled. When spectre mitigations are enabled then the offending shape of load is not generated. When signals-based-traps are disabled then spectre mitigations are also automatically disabled. The specific bug in Cranelift is a miscompile of a load of the shape load(iadd(base, ishl(index, amt))) where amt is a constant. The amt value is masked incorrectly to test if it's a certain value, and this incorrect mask means that Cranelift can pattern-match this lowering rule during instruction selection erroneously, diverging from WebAssembly's and Cranelift's semantics. This incorrect lowering would, for example, load an address much further away than intended as the correct address's computation would have wrapped around to a smaller value insetad. This vulnerability is fixed in 36.0.7, 42.0.2, and 43.0.1.

An out-of-bound memory read vulnerability was discovered in Bytecode Alliance wasm-micro-runtime v2.0.0 which allows a remote attacker to cause a denial of service via the "block_type_get_arity" function in core/iwasm/interpreter/wasm.h.

Wasmtime is a runtime for WebAssembly. From 25.0.0 to before 36.0.7, 42.0.2, and 43.0.1, Wasmtime with its Winch (baseline) non-default compiler backend may allow properly constructed guest Wasm to access host memory outside of its linear-memory sandbox. This vulnerability requires use of the Winch compiler (-Ccompiler=winch). By default, Wasmtime uses its Cranelift backend, not Winch. With Winch, the same incorrect assumption is present in theory on both aarch64 and x86-64. The aarch64 case has an observed-working proof of concept, while the x86-64 case is theoretical and may not be reachable in practice. This Winch compiler bug can allow the Wasm guest to access memory before or after the linear-memory region, independently of whether pre- or post-guard regions are configured. The accessible range in the initial bug proof-of-concept is up to 32KiB before the start of memory, or ~4GiB after the start of memory, independently of the size of pre- or post-guard regions or the use of explicit or guard-region-based bounds checking. However, the underlying bug assumes a 32-bit memory offset stored in a 64-bit register has its upper bits cleared when it may not, and so closely related variants of the initial proof-of-concept may be able to access truly arbitrary memory in-process. This could result in a host process segmentation fault (DoS), an arbitrary data leak from the host process, or with a write, potentially an arbitrary RCE. This vulnerability is fixed in 36.0.7, 42.0.2, and 43.0.1.

Wasmtime is a runtime for WebAssembly. Starting in version 29.0.0 and prior to version 36.0.5, 40.0.3, and 41.0.1, on x86-64 platforms with AVX, Wasmtime's compilation of the `f64.copysign` WebAssembly instruction with Cranelift may load 8 more bytes than is necessary. When signals-based-traps are disabled this can result in a uncaught segfault due to loading from unmapped guard pages. With guard pages disabled it's possible for out-of-sandbox data to be loaded, but unless there is another bug in Cranelift this data is not visible to WebAssembly guests. Wasmtime 36.0.5, 40.0.3, and 41.0.1 have been released to fix this issue. Users are recommended to upgrade to the patched versions of Wasmtime. Other affected versions are not patched and users should updated to supported major version instead. This bug can be worked around by enabling signals-based-traps. While disabling guard pages can be a quick fix in some situations, it's not recommended to disabled guard pages as it is a key defense-in-depth measure of Wasmtime.

wasmtime is a fast and secure runtime for WebAssembly. In affected versions wasmtime's code generator, Cranelift, has a bug on x86_64 targets where address-mode computation mistakenly would calculate a 35-bit effective address instead of WebAssembly's defined 33-bit effective address. This bug means that, with default codegen settings, a wasm-controlled load/store operation could read/write addresses up to 35 bits away from the base of linear memory. Due to this bug, however, addresses up to `0xffffffff * 8 + 0x7ffffffc = 36507222004 = ~34G` bytes away from the base of linear memory are possible from guest code. This means that the virtual memory 6G away from the base of linear memory up to ~34G away can be read/written by a malicious module. A guest module can, without the knowledge of the embedder, read/write memory in this region. The memory may belong to other WebAssembly instances when using the pooling allocator, for example. Affected embedders are recommended to analyze preexisting wasm modules to see if they're affected by the incorrect codegen rules and possibly correlate that with an anomalous number of traps during historical execution to locate possibly suspicious modules. The specific bug in Cranelift's x86_64 backend is that a WebAssembly address which is left-shifted by a constant amount from 1 to 3 will get folded into x86_64's addressing modes which perform shifts. For example `(i32.load (i32.shl (local.get 0) (i32.const 3)))` loads from the WebAssembly address `$local0 << 3`. When translated to Cranelift the `$local0 << 3` computation, a 32-bit value, is zero-extended to a 64-bit value and then added to the base address of linear memory. Cranelift would generate an instruction of the form `movl (%base, %local0, 8), %dst` which calculates `%base + %local0 << 3`. The bug here, however, is that the address computation happens with 64-bit values, where the `$local0 << 3` computation was supposed to be truncated to a a 32-bit value. This means that `%local0`, which can use up to 32-bits for an address, gets 3 extra bits of address space to be accessible via this `movl` instruction. The fix in Cranelift is to remove the erroneous lowering rules in the backend which handle these zero-extended expression. The above example is then translated to `movl %local0, %temp; shl $3, %temp; movl (%base, %temp), %dst` which correctly truncates the intermediate computation of `%local0 << 3` to 32-bits inside the `%temp` register which is then added to the `%base` value. Wasmtime version 4.0.1, 5.0.1, and 6.0.1 have been released and have all been patched to no longer contain the erroneous lowering rules. While updating Wasmtime is recommended, there are a number of possible workarounds that embedders can employ to mitigate this issue if updating is not possible. Note that none of these workarounds are on-by-default and require explicit configuration: 1. The `Config::static_memory_maximum_size(0)` option can be used to force all accesses to linear memory to be explicitly bounds-checked. This will perform a bounds check separately from the address-mode computation which correctly calculates the effective address of a load/store. Note that this can have a large impact on the execution performance of WebAssembly modules. 2. The `Config::static_memory_guard_size(1 << 36)` option can be used to greatly increase the guard pages placed after linear memory. This will guarantee that memory accesses up-to-34G away are guaranteed to be semantically correct by reserving unmapped memory for the instance. Note that this reserves a very large amount of virtual memory per-instances and can greatly reduce the maximum number of concurrent instances being run. 3. If using a non-x86_64 host is possible, then that will also work around this bug. This bug does not affect Wasmtime's or Cranelift's AArch64 backend, for example.

Wasmtime is a standalone runtime for WebAssembly. Prior to version 2.0.2, there is a bug in Wasmtime's implementation of its pooling instance allocator when the allocator is configured to give WebAssembly instances a maximum of zero pages of memory. In this configuration, the virtual memory mapping for WebAssembly memories did not meet the compiler-required configuration requirements for safely executing WebAssembly modules. Wasmtime's default settings require virtual memory page faults to indicate that wasm reads/writes are out-of-bounds, but the pooling allocator's configuration would not create an appropriate virtual memory mapping for this meaning out of bounds reads/writes can successfully read/write memory unrelated to the wasm sandbox within range of the base address of the memory mapping created by the pooling allocator. This bug is not applicable with the default settings of the `wasmtime` crate. This bug can only be triggered by setting `InstanceLimits::memory_pages` to zero. This is expected to be a very rare configuration since this means that wasm modules cannot allocate any pages of linear memory. All wasm modules produced by all current toolchains are highly likely to use linear memory, so it's expected to be unlikely that this configuration is set to zero by any production embedding of Wasmtime. This bug has been patched and users should upgrade to Wasmtime 2.0.2. This bug can be worked around by increasing the `memory_pages` allotment when configuring the pooling allocator to a value greater than zero. If an embedding wishes to still prevent memory from actually being used then the `Store::limiter` method can be used to dynamically disallow growth of memory beyond 0 bytes large. Note that the default `memory_pages` value is greater than zero.

In nci_proc_rf_management_ntf of nci_hrcv.cc, there is a possible out of bounds read due to a missing bounds check. This could lead to local escalation of privilege with System execution privileges needed. User interaction is not needed for exploitation.Product: AndroidVersions: Android-11Android ID: A-164440989

Out of bounds read in the firmware for some Intel(R) Processors may allow an authenticated user to potentially enable escalation of privilege via local access.

HUAWEI nova 4 versions earlier than 10.0.0.165(C01E34R2P4) and SydneyM-AL00 versions earlier than 10.0.0.165(C00E66R1P5) have an out-of-bounds read and write vulnerability. An attacker with specific permissions crafts malformed packet with specific parameter and sends the packet to the affected products. Due to insufficient validation of packet, which may be exploited to cause the information leakage or arbitrary code execution.

An issue was discovered on Samsung mobile devices with P(9.0) (Exynos chipsets) software. Kernel Wi-Fi drivers allow out-of-bounds Read or Write operations (e.g., a buffer overflow). The Samsung IDs are SVE-2019-16125, SVE-2019-16134, SVE-2019-16158, SVE-2019-16159, SVE-2019-16319, SVE-2019-16320, SVE-2019-16337, SVE-2019-16464, SVE-2019-16465, SVE-2019-16467 (March 2020).

Out of bound read in BIOS firmware for 8th, 9th Generation Intel(R) Core(TM), Intel(R) Celeron(R) Processor 4000 Series Processors may allow an unauthenticated user to potentially enable elevation of privilege or denial of service via local access.

Out of bounds read in some Intel(R) Graphics Drivers before versions 15.45.31.5127 and 15.40.45.5126 may allow an authenticated user to potentially enable escalation of privilege via local access.

In skb_headlen of /include/linux/skbuff.h, there is a possible out of bounds read due to memory corruption. This could lead to local escalation of privilege with no additional execution privileges needed. User interaction is not needed for exploitation.Product: AndroidVersions: Android kernelAndroid ID: A-153881554

NVIDIA CUDA Toolkit, all versions prior to 11.1.1, contains a vulnerability in the NVJPEG library in which an out-of-bounds read or write operation may lead to code execution, denial of service, or information disclosure.

OpenWrt Project is a Linux operating system targeting embedded devices. Prior to version 24.10.4, local users could read and write arbitrary kernel memory using the ioctls of the ltq-ptm driver which is used to drive the datapath of the DSL line. This only effects the lantiq target supporting xrx200, danube and amazon SoCs from Lantiq/Intel/MaxLinear with the DSL in PTM mode. The DSL driver for the VRX518 is not affected. ATM mode is also not affected. Most VDSL lines use PTM mode and most ADSL lines use ATM mode. OpenWrt is normally running as a single user system, but some services are sandboxed. This vulnerability could allow attackers to escape a ujail sandbox or other contains. This is fixed in OpenWrt 24.10.4. There are no workarounds.

Improper casting of structure while handling the buffer leads to out of bound read in display in Snapdragon Auto, Snapdragon Connectivity, Snapdragon Consumer IOT, Snapdragon Industrial IOT, Snapdragon IoT, Snapdragon Mobile, Snapdragon Voice & Music, Snapdragon Wearables in MDM9150, MDM9206, MDM9607, MDM9650, MSM8909W, MSM8996AU, QCS405, QCS605, Qualcomm 215, SD 210/SD 212/SD 205, SD 425, SD 427, SD 430, SD 435, SD 439 / SD 429, SD 450, SD 615/16/SD 415, SD 625, SD 632, SD 636, SD 665, SD 675, SD 712 / SD 710 / SD 670, SD 730, SD 820, SD 820A, SD 835, SD 845 / SD 850, SD 855, SDA660, SDM439, SDM630, SDM660, SDX20

Possibility of out-of-bound read if id received from SPI is not in range of FIFO in Snapdragon Auto, Snapdragon Compute, Snapdragon Consumer IOT, Snapdragon Industrial IOT, Snapdragon Mobile, Snapdragon Wearables, Snapdragon Wired Infrastructure and Networking in IPQ4019, IPQ8064, MSM8909W, MSM8996AU, QCA9980, QCS605, Qualcomm 215, SD 425, SD 439 / SD 429, SD 450, SD 625, SD 632, SD 636, SD 712 / SD 710 / SD 670, SD 820A, SD 845 / SD 850, SD 855, SDM439, SDM660, SDX24

Out of bound read can happen due to lack of NULL termination on user controlled data in WLAN in Snapdragon Auto, Snapdragon Compute, Snapdragon Consumer IOT, Snapdragon Industrial IOT, Snapdragon Mobile, Snapdragon Voice & Music in MSM8996AU, QCS405, QCS605, SD 210/SD 212/SD 205, SD 425, SD 427, SD 430, SD 435, SD 450, SD 625, SD 636, SD 665, SD 675, SD 712 / SD 710 / SD 670, SD 730, SD 820A, SD 835, SD 845 / SD 850, SD 855, SDA660, SDM630, SDM660, SDX24

vcs_write in drivers/tty/vt/vc_screen.c in the Linux kernel through 5.3.13 does not prevent write access to vcsu devices, aka CID-0c9acb1af77a.

There is a heap-based buffer over-read in the _nc_find_entry function in tinfo/comp_hash.c in the terminfo library in ncurses before 6.1-20191012.

Tensorflow is an Open Source Machine Learning Framework. The TFG dialect of TensorFlow (MLIR) makes several assumptions about the incoming `GraphDef` before converting it to the MLIR-based dialect. If an attacker changes the `SavedModel` format on disk to invalidate these assumptions and the `GraphDef` is then converted to MLIR-based IR then they can cause a crash in the Python interpreter. Under certain scenarios, heap OOB read/writes are possible. These issues have been discovered via fuzzing and it is possible that more weaknesses exist. We will patch them as they are discovered.

An issue was discovered in drivers/scsi/qedi/qedi_dbg.c in the Linux kernel before 5.1.12. In the qedi_dbg_* family of functions, there is an out-of-bounds read.

In the Linux kernel before 5.2.3, set_geometry in drivers/block/floppy.c does not validate the sect and head fields, as demonstrated by an integer overflow and out-of-bounds read. It can be triggered by an unprivileged local user when a floppy disk has been inserted. NOTE: QEMU creates the floppy device by default.

Possible out of bound memory access while playing a crafted clip in media player in Snapdragon Auto, Snapdragon Compute, Snapdragon Consumer IOT, Snapdragon Industrial IOT, Snapdragon Mobile, Snapdragon Voice & Music, Snapdragon Wearables in SM8150

Buffer Over-read in GitHub repository bfabiszewski/libmobi prior to 0.11. This vulnerability is capable of arbitrary code execution.

Buffer Over-read in GitHub repository vim/vim prior to 8.2.4974.

Out-of-bounds Read in mrb_obj_is_kind_of in in GitHub repository mruby/mruby prior to 3.2. # Impact: Possible arbitrary code execution if being exploited.

In getAppSize of InstalldNativeService.cpp, there is a possible out of bounds read due to a missing bounds check. This could lead to local escalation of privilege with System execution privileges needed. User interaction is not needed for exploitation.Product: AndroidVersions: Android-12LAndroid ID: A-220733817

In ccci, there is a possible out of bounds read due to a missing bounds check. This could lead to local information disclosure with System execution privileges needed. User interaction is not needed for exploitation. Patch ID: ALPS06108658; Issue ID: ALPS06108658.

Out-of-bounds read in Microsoft Input Method Editor (IME) allows an authorized attacker to elevate privileges locally.

In __f2fs_setxattr in fs/f2fs/xattr.c in the Linux kernel through 5.15.11, there is an out-of-bounds memory access when an inode has an invalid last xattr entry.

Possible OOB issue in EEPROM due to lack of check while accessing memory map array at the time of reading operation in Snapdragon Auto, Snapdragon Compute, Snapdragon Consumer Electronics Connectivity, Snapdragon Consumer IOT, Snapdragon Industrial IOT, Snapdragon Mobile, Snapdragon Voice & Music, Snapdragon Wearables in APQ8009, APQ8053, MSM8909W, MSM8917, MSM8953, Nicobar, QCS405, QCS605, QM215, SA6155P, SDA845, SDM429, SDM439, SDM450, SDM632, SDM670, SDM710, SDM845, SDX24, SDX55, SM6150, SM7150, SM8150, SM8250, SXR1130, SXR2130

Out of bound read/write in system firmware for Intel(R) NUC Kit may allow a privileged user to potentially enable escalation of privilege, denial of service and/or information disclosure via local access.

Possibility of out of bound access in debug queue, if packet size field is corrupted in Snapdragon Auto, Snapdragon Compute, Snapdragon Connectivity, Snapdragon Consumer Electronics Connectivity, Snapdragon Consumer IOT, Snapdragon Industrial IOT, Snapdragon Mobile, Snapdragon Voice & Music, Snapdragon Wearables in APQ8009, APQ8017, APQ8053, APQ8096AU, APQ8098, MDM9206, MDM9207C, MDM9607, MDM9640, MDM9650, MSM8909W, MSM8917, MSM8920, MSM8937, MSM8940, MSM8953, MSM8996AU, MSM8998, Nicobar, QCN7605, QCS405, QCS605, QM215, SA6155P, SDA660, SDA845, SDM429, SDM439, SDM450, SDM630, SDM632, SDM636, SDM660, SDM670, SDM710, SDM845, SDX20, SDX24, SDX55, SM6150, SM7150, SM8150, SM8250, SXR1130, SXR2130

Lack of check of extscan change results received from firmware can lead to an out of buffer read in Snapdragon Auto, Snapdragon Consumer Electronics Connectivity, Snapdragon Consumer IOT, Snapdragon Industrial IOT, Snapdragon Mobile, Snapdragon Voice & Music in MDM9150, MDM9206, MDM9607, MDM9640, MDM9650, MSM8996AU, QCA6174A, QCA6574AU, QCA9377, QCA9379, QCS605, SD 210/SD 212/SD 205, SD 425, SD 430, SD 600, SD 625, SD 636, SD 665, SD 675, SD 712 / SD 710 / SD 670, SD 730, SD 820A, SD 835, SD 845 / SD 850, SD 855, SDA660, SDM630, SDM660, SDX20, SDX24

While handling the vendor command there is an integer truncation issue that could yield a buffer overflow due to int data type copied to u8 data type in Snapdragon Auto, Snapdragon Compute, Snapdragon Consumer Electronics Connectivity, Snapdragon Industrial IOT, Snapdragon Mobile in APQ8096AU, MSM8996AU, QCA6574AU, QCN7605, Rennell, SC8180X, SDM710, SDX55, SM7150, SM8150, SM8250, SXR2130

TensorFlow is an open source platform for machine learning. In affected versions the code for sparse matrix multiplication is vulnerable to undefined behavior via binding a reference to `nullptr`. This occurs whenever the dimensions of `a` or `b` are 0 or less. In the case on one of these is 0, an empty output tensor should be allocated (to conserve the invariant that output tensors are always allocated when the operation is successful) but nothing should be written to it (that is, we should return early from the kernel implementation). Otherwise, attempts to write to this empty tensor would result in heap OOB access. The fix will be included in TensorFlow 2.7.0. We will also cherrypick this commit on TensorFlow 2.6.1, TensorFlow 2.5.2, and TensorFlow 2.4.4, as these are also affected and still in supported range.

TensorFlow is an end-to-end open source platform for machine learning. In affected versions an attacker can cause undefined behavior via binding a reference to null pointer in all binary cwise operations that don't require broadcasting (e.g., gradients of binary cwise operations). The [implementation](https://github.com/tensorflow/tensorflow/blob/84d053187cb80d975ef2b9684d4b61981bca0c41/tensorflow/core/kernels/cwise_ops_common.h#L264) assumes that the two inputs have exactly the same number of elements but does not check that. Hence, when the eigen functor executes it triggers heap OOB reads and undefined behavior due to binding to nullptr. We have patched the issue in GitHub commit 93f428fd1768df147171ed674fee1fc5ab8309ec. The fix will be included in TensorFlow 2.6.0. We will also cherrypick this commit on TensorFlow 2.5.1, TensorFlow 2.4.3, and TensorFlow 2.3.4, as these are also affected and still in supported range.

TensorFlow is an end-to-end open source platform for machine learning. When restoring tensors via raw APIs, if the tensor name is not provided, TensorFlow can be tricked into dereferencing a null pointer. Alternatively, attackers can read memory outside the bounds of heap allocated data by providing some tensor names but not enough for a successful restoration. The [implementation](https://github.com/tensorflow/tensorflow/blob/47a06f40411a69c99f381495f490536972152ac0/tensorflow/core/kernels/save_restore_tensor.cc#L158-L159) retrieves the tensor list corresponding to the `tensor_name` user controlled input and immediately retrieves the tensor at the restoration index (controlled via `preferred_shard` argument). This occurs without validating that the provided list has enough values. If the list is empty this results in dereferencing a null pointer (undefined behavior). If, however, the list has some elements, if the restoration index is outside the bounds this results in heap OOB read. We have patched the issue in GitHub commit 9e82dce6e6bd1f36a57e08fa85af213e2b2f2622. The fix will be included in TensorFlow 2.6.0. We will also cherrypick this commit on TensorFlow 2.5.1, TensorFlow 2.4.3, and TensorFlow 2.3.4, as these are also affected and still in supported range.