In the Linux kernel, the following vulnerability has been resolved: f2fs: fix fsck inconsistency caused by FGGC of node block During FGGC node block migration, fsck may incorrectly treat the migrated node block as fsync-written data. The reproduction scenario: root@vm:/mnt/f2fs# seq 1 2048 | xargs -n 1 ./test_sync // write inline inode and sync root@vm:/mnt/f2fs# rm -f 1 root@vm:/mnt/f2fs# sync root@vm:/mnt/f2fs# f2fs_io gc_range // move data block in sync mode and not write CP SPO, "fsck --dry-run" find inode has already checkpointed but still with DENT_BIT_SHIFT set The root cause is that GC does not clear the dentry mark and fsync mark during node block migration, leading fsck to misinterpret them as user-issued fsync writes. In BGGC mode, node block migration is handled by f2fs_sync_node_pages(), which guarantees the dentry and fsync marks are cleared before writing. This patch move the set/clear of the fsync|dentry marks into __write_node_folio to make the logic clearer, and ensures the fsync|dentry mark is cleared in FGGC.

linux

Aegra is a drop-in replacement for LangSmith Deployments. Prior to 0.9.7, with multiple authenticated users on a shared instance are vulnerable to a cross-tenant IDOR. Any authenticated attacker, given another user's thread_id, can execute graph runs against the user's thread, read the user's full checkpoint state, and inject arbitrary messages into the user's conversation history. This vulnerability is fixed in 0.9.7.

The snorkel library thru v0.10.0 contains an insecure deserialization vulnerability (CWE-502) in the Trainer.load() method of the Trainer class. The method loads model checkpoint files using torch.load() without enabling the security-restrictive weights_only=True parameter. This default behavior allows the deserialization of arbitrary Python objects via the Pickle module. A remote attacker can exploit this by providing a maliciously crafted model file, leading to arbitrary code execution on the victim's system when the file is loaded via the vulnerable method.

snorkel

PyTorch-Lightning versions 2.6.0 and earlier contain an insecure deserialization vulnerability (CWE-502) in the checkpoint loading mechanism. The LightningModule.load_from_checkpoint() method, which is commonly used to load saved model states, internally calls torch.load() without setting the security-restrictive weights_only=True parameter. This default behavior allows the deserialization of arbitrary Python objects via the Pickle module. A remote attacker can exploit this by providing a maliciously crafted checkpoint file, leading to arbitrary code execution on the victim's system when the file is loaded.

lightningai

The torch-checkpoint-shrink.py script in the ml-engineering project in commit 0099885db36a8f06556efe1faf552518852cb1e0 (2025-20-27) contains an insecure deserialization vulnerability (CWE-502). The script uses torch.load() to process PyTorch checkpoint files (.pt) without enabling the security-restrictive weights_only=True parameter. This oversight allows the deserialization of arbitrary Python objects via the pickle module. A remote attacker can exploit this by providing a maliciously crafted checkpoint file, leading to arbitrary code execution in the context of the user running the script.

The flash-attention training framework thru commit e724e2588cbe754beb97cf7c011b5e7e34119e62 (2025-13-04) contains an insecure deserialization vulnerability (CWE-502) in its checkpoint loading mechanism. The load_checkpoint() function in checkpoint.py and the checkpoint loading code in eval.py use torch.load() without enabling the security-restrictive weights_only=True parameter. This allows the deserialization of arbitrary Python objects via the pickle module. An attacker can exploit this by providing a maliciously crafted checkpoint file. When a victim loads this checkpoint during model warmstarting or evaluation, arbitrary code is executed on the victim's system.

CosyVoice thru commit 6e01309e01bc93bbeb83bdd996b1182a81aaf11e (2025-30-21) contains an insecure deserialization vulnerability (CWE-502) in its average_model.py model averaging tool. The script loads PyTorch checkpoint files (epoch_*.pt) for model averaging using torch.load() without enabling the weights_only=True security parameter. This allows the deserialization of arbitrary Python objects via the pickle module. An attacker can exploit this by providing malicious checkpoint files within a directory. When a victim uses the tool to average models from this directory, arbitrary code is executed on the victim's system.

In the Linux kernel, the following vulnerability has been resolved: ALSA: mixer: oss: Add card disconnect checkpoints ALSA OSS mixer layer calls the kcontrol ops rather individually, and pending calls might be not always caught at disconnecting the device. For avoiding the potential UAF scenarios, add sanity checks of the card disconnection at each entry point of OSS mixer accesses. The rwsem is taken just before that check, hence the rest context should be covered by that properly.

linux

In the Linux kernel, the following vulnerability has been resolved: btrfs: reject root items with drop_progress and zero drop_level [BUG] When recovering relocation at mount time, merge_reloc_root() and btrfs_drop_snapshot() both use BUG_ON(level == 0) to guard against an impossible state: a non-zero drop_progress combined with a zero drop_level in a root_item, which can be triggered: ------------[ cut here ]------------ kernel BUG at fs/btrfs/relocation.c:1545! Oops: invalid opcode: 0000 [#1] SMP KASAN NOPTI CPU: 1 UID: 0 PID: 283 ... Tainted: 6.18.0+ #16 PREEMPT(voluntary) Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE Hardware name: QEMU Ubuntu 24.04 PC v2, BIOS 1.16.3-debian-1.16.3-2 RIP: 0010:merge_reloc_root+0x1266/0x1650 fs/btrfs/relocation.c:1545 Code: ffff0000 00004589 d7e9acfa ffffe8a1 79bafebe 02000000 Call Trace: merge_reloc_roots+0x295/0x890 fs/btrfs/relocation.c:1861 btrfs_recover_relocation+0xd6e/0x11d0 fs/btrfs/relocation.c:4195 btrfs_start_pre_rw_mount+0xa4d/0x1810 fs/btrfs/disk-io.c:3130 open_ctree+0x5824/0x5fe0 fs/btrfs/disk-io.c:3640 btrfs_fill_super fs/btrfs/super.c:987 [inline] btrfs_get_tree_super fs/btrfs/super.c:1951 [inline] btrfs_get_tree_subvol fs/btrfs/super.c:2094 [inline] btrfs_get_tree+0x111c/0x2190 fs/btrfs/super.c:2128 vfs_get_tree+0x9a/0x370 fs/super.c:1758 fc_mount fs/namespace.c:1199 [inline] do_new_mount_fc fs/namespace.c:3642 [inline] do_new_mount fs/namespace.c:3718 [inline] path_mount+0x5b8/0x1ea0 fs/namespace.c:4028 do_mount fs/namespace.c:4041 [inline] __do_sys_mount fs/namespace.c:4229 [inline] __se_sys_mount fs/namespace.c:4206 [inline] __x64_sys_mount+0x282/0x320 fs/namespace.c:4206 ... RIP: 0033:0x7f969c9a8fde Code: 0f1f4000 48c7c2b0 fffffff7 d8648902 b8ffffff ffc3660f ---[ end trace 0000000000000000 ]--- The bug is reproducible on 7.0.0-rc2-next-20260310 with our dynamic metadata fuzzing tool that corrupts btrfs metadata at runtime. [CAUSE] A non-zero drop_progress.objectid means an interrupted btrfs_drop_snapshot() left a resume point on disk, and in that case drop_level must be greater than 0 because the checkpoint is only saved at internal node levels. Although this invariant is enforced when the kernel writes the root item, it is not validated when the root item is read back from disk. That allows on-disk corruption to provide an invalid state with drop_progress.objectid != 0 and drop_level == 0. When relocation recovery later processes such a root item, merge_reloc_root() reads drop_level and hits BUG_ON(level == 0). The same invalid metadata can also trigger the corresponding BUG_ON() in btrfs_drop_snapshot(). [FIX] Fix this by validating the root_item invariant in tree-checker when reading root items from disk: if drop_progress.objectid is non-zero, drop_level must also be non-zero. Reject such malformed metadata with -EUCLEAN before it reaches merge_reloc_root() or btrfs_drop_snapshot() and triggers the BUG_ON. After the fix, the same corruption is correctly rejected by tree-checker and the BUG_ON is no longer triggered.

linux

In the Linux kernel, the following vulnerability has been resolved: f2fs: fix UAF caused by decrementing sbi->nr_pages[] in f2fs_write_end_io() The xfstests case "generic/107" and syzbot have both reported a NULL pointer dereference. The concurrent scenario that triggers the panic is as follows: F2FS_WB_CP_DATA write callback umount - f2fs_write_checkpoint - f2fs_wait_on_all_pages(sbi, F2FS_WB_CP_DATA) - blk_mq_end_request - bio_endio - f2fs_write_end_io : dec_page_count(sbi, F2FS_WB_CP_DATA) : wake_up(&sbi->cp_wait) - kill_f2fs_super - kill_block_super - f2fs_put_super : iput(sbi->node_inode) : sbi->node_inode = NULL : f2fs_in_warm_node_list - is_node_folio // sbi->node_inode is NULL and panic The root cause is that f2fs_put_super() calls iput(sbi->node_inode) and sets sbi->node_inode to NULL after sbi->nr_pages[F2FS_WB_CP_DATA] is decremented to zero. As a result, f2fs_in_warm_node_list() may dereference a NULL node_inode when checking whether a folio belongs to the node inode, leading to a panic. This patch fixes the issue by calling f2fs_in_warm_node_list() before decrementing sbi->nr_pages[F2FS_WB_CP_DATA], thus preventing the use-after-free condition.

linux

A vulnerability in the HuggingFace Transformers library, specifically in the `Trainer` class, allows for arbitrary code execution. The `_load_rng_state()` method in `src/transformers/trainer.py` at line 3059 calls `torch.load()` without the `weights_only=True` parameter. This issue affects all versions of the library supporting `torch>=2.2` when used with PyTorch versions below 2.6, as the `safe_globals()` context manager provides no protection in these versions. An attacker can exploit this vulnerability by supplying a malicious checkpoint file, such as `rng_state.pth`, which can execute arbitrary code when loaded. The issue is resolved in version v5.0.0rc3.

huggingface

A directory traversal vulnerability in the agentic-context-engine project versions up to 0.7.1 allows arbitrary file writes via the checkpoint_dir parameter in OfflineACE.run. The save_to_file method in ace/skillbook.py fails to normalize or validate filesystem paths, allowing traversal sequences to escape the intended checkpoint directory. This vulnerability allows attackers to overwrite arbitrary files accessible to the application process, potentially leading to application corruption, privilege escalation, or code execution depending on the deployment context.

NVIDIA NeMo Framework contains a vulnerability in checkpoint loading where an attacker could cause remote code execution. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure and data tampering.

nvidia

NVIDIA Megatron-LM contains a vulnerability in checkpoint loading where an Attacker may cause an RCE by convincing a user to load a maliciously crafted file. A successful exploit of this vulnerability may lead to code execution, escalation of privileges, information disclosure, and data tampering.

nvidia

NVIDIA Megatron-LM contains a vulnerability in checkpoint loading where an Attacker may cause an RCE by convincing a user to load a maliciously crafted file. A successful exploit of this vulnerability may lead to code execution, escalation of privileges, information disclosure, and data tampering.

nvidia

In the Linux kernel, the following vulnerability has been resolved: f2fs: fix IS_CHECKPOINTED flag inconsistency issue caused by concurrent atomic commit and checkpoint writes During SPO tests, when mounting F2FS, an -EINVAL error was returned from f2fs_recover_inode_page. The issue occurred under the following scenario Thread A Thread B f2fs_ioc_commit_atomic_write - f2fs_do_sync_file // atomic = true - f2fs_fsync_node_pages : last_folio = inode folio : schedule before folio_lock(last_folio) f2fs_write_checkpoint - block_operations// writeback last_folio - schedule before f2fs_flush_nat_entries : set_fsync_mark(last_folio, 1) : set_dentry_mark(last_folio, 1) : folio_mark_dirty(last_folio) - __write_node_folio(last_folio) : f2fs_down_read(&sbi->node_write)//block - f2fs_flush_nat_entries : {struct nat_entry}->flag |= BIT(IS_CHECKPOINTED) - unblock_operations : f2fs_up_write(&sbi->node_write) f2fs_write_checkpoint//return : f2fs_do_write_node_page() f2fs_ioc_commit_atomic_write//return SPO Thread A calls f2fs_need_dentry_mark(sbi, ino), and the last_folio has already been written once. However, the {struct nat_entry}->flag did not have the IS_CHECKPOINTED set, causing set_dentry_mark(last_folio, 1) and write last_folio again after Thread B finishes f2fs_write_checkpoint. After SPO and reboot, it was detected that {struct node_info}->blk_addr was not NULL_ADDR because Thread B successfully write the checkpoint. This issue only occurs in atomic write scenarios. For regular file fsync operations, the folio must be dirty. If block_operations->f2fs_sync_node_pages successfully submit the folio write, this path will not be executed. Otherwise, the f2fs_write_checkpoint will need to wait for the folio write submission to complete, as sbi->nr_pages[F2FS_DIRTY_NODES] > 0. Therefore, the situation where f2fs_need_dentry_mark checks that the {struct nat_entry}->flag /wo the IS_CHECKPOINTED flag, but the folio write has already been submitted, will not occur. Therefore, for atomic file fsync, sbi->node_write should be acquired through __write_node_folio to ensure that the IS_CHECKPOINTED flag correctly indicates that the checkpoint write has been completed.

linux

LangGraph SQLite Checkpoint is an implementation of LangGraph CheckpointSaver that uses SQLite DB (both sync and async, via aiosqlite). In version 1.0.9 and prior, LangGraph checkpointers can load msgpack-encoded checkpoints that reconstruct Python objects during deserialization. If an attacker can modify checkpoint data in the backing store (for example, after a database compromise or other privileged write access to the persistence layer), they can potentially supply a crafted payload that triggers unsafe object reconstruction when the checkpoint is loaded. No known patch is public.

langchain

In the Linux kernel, the following vulnerability has been resolved: Revert "f2fs: block cache/dio write during f2fs_enable_checkpoint()" This reverts commit 196c81fdd438f7ac429d5639090a9816abb9760a. Original patch may cause below deadlock, revert it. write remount - write_begin - lock_page --- lock A - prepare_write_begin - f2fs_map_lock - f2fs_enable_checkpoint - down_write(cp_enable_rwsem) --- lock B - sync_inode_sb - writepages - lock_page --- lock A - down_read(cp_enable_rwsem) --- lock A

linux

LangGraph Checkpoint defines the base interface for LangGraph checkpointers. Prior to version 4.0.0, a Remote Code Execution vulnerability exists in LangGraph's caching layer when applications enable cache backends that inherit from `BaseCache` and opt nodes into caching via `CachePolicy`. Prior to `langgraph-checkpoint` 4.0.0, `BaseCache` defaults to `JsonPlusSerializer(pickle_fallback=True)`. When msgpack serialization fails, cached values can be deserialized via `pickle.loads(...)`. Caching is not enabled by default. Applications are affected only when the application explicitly enables a cache backend (for example by passing `cache=...` to `StateGraph.compile(...)` or otherwise configuring a `BaseCache` implementation), one or more nodes opt into caching via `CachePolicy`, and the attacker can write to the cache backend (for example a network-accessible Redis instance with weak/no auth, shared cache infrastructure reachable by other tenants/services, or a writable SQLite cache file). An attacker must be able to write attacker-controlled bytes into the cache backend such that the LangGraph process later reads and deserializes them. This typically requires write access to a networked cache (for example a network-accessible Redis instance with weak/no auth or shared cache infrastructure reachable by other tenants/services) or write access to local cache storage (for example a writable SQLite cache file via permissive file permissions or a shared writable volume). Because exploitation requires write access to the cache storage layer, this is a post-compromise / post-access escalation vector. LangGraph Checkpoint 4.0.0 patches the issue.

@langchain/langgraph-checkpoint-redis is the Redis checkpoint and store implementation for LangGraph. A query injection vulnerability exists in the @langchain/langgraph-checkpoint-redis package's filter handling. The RedisSaver and ShallowRedisSaver classes construct RediSearch queries by directly interpolating user-provided filter keys and values without proper escaping. RediSearch has special syntax characters that can modify query behavior, and when user-controlled data contains these characters, the query logic can be manipulated to bypass intended access controls. This vulnerability is fixed in 1.0.2.

PyTorch is a Python package that provides tensor computation. Prior to version 2.10.0, a vulnerability in PyTorch's `weights_only` unpickler allows an attacker to craft a malicious checkpoint file (`.pth`) that, when loaded with `torch.load(..., weights_only=True)`, can corrupt memory and potentially lead to arbitrary code execution. Version 2.10.0 fixes the issue.

linuxfoundation

In the Linux kernel, the following vulnerability has been resolved: f2fs: fix return value of f2fs_recover_fsync_data() With below scripts, it will trigger panic in f2fs: mkfs.f2fs -f /dev/vdd mount /dev/vdd /mnt/f2fs touch /mnt/f2fs/foo sync echo 111 >> /mnt/f2fs/foo f2fs_io fsync /mnt/f2fs/foo f2fs_io shutdown 2 /mnt/f2fs umount /mnt/f2fs mount -o ro,norecovery /dev/vdd /mnt/f2fs or mount -o ro,disable_roll_forward /dev/vdd /mnt/f2fs F2FS-fs (vdd): f2fs_recover_fsync_data: recovery fsync data, check_only: 0 F2FS-fs (vdd): Mounted with checkpoint version = 7f5c361f F2FS-fs (vdd): Stopped filesystem due to reason: 0 F2FS-fs (vdd): f2fs_recover_fsync_data: recovery fsync data, check_only: 1 Filesystem f2fs get_tree() didn't set fc->root, returned 1 ------------[ cut here ]------------ kernel BUG at fs/super.c:1761! Oops: invalid opcode: 0000 [#1] SMP PTI CPU: 3 UID: 0 PID: 722 Comm: mount Not tainted 6.18.0-rc2+ #721 PREEMPT(voluntary) Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014 RIP: 0010:vfs_get_tree.cold+0x18/0x1a Call Trace: <TASK> fc_mount+0x13/0xa0 path_mount+0x34e/0xc50 __x64_sys_mount+0x121/0x150 do_syscall_64+0x84/0x800 entry_SYSCALL_64_after_hwframe+0x76/0x7e RIP: 0033:0x7fa6cc126cfe The root cause is we missed to handle error number returned from f2fs_recover_fsync_data() when mounting image w/ ro,norecovery or ro,disable_roll_forward mount option, result in returning a positive error number to vfs_get_tree(), fix it.

LMDeploy is a toolkit for compressing, deploying, and serving LLMs. Prior to version 0.11.1, an insecure deserialization vulnerability exists in lmdeploy where torch.load() is called without the weights_only=True parameter when loading model checkpoint files. This allows an attacker to execute arbitrary code on the victim's machine when they load a malicious .bin or .pt model file. This issue has been patched in version 0.11.1.

internlm