Search: t**n*e**l s.a. c***t**c***a

In the Linux kernel, the following vulnerability has been resolved: drm/xe/rtp: Add RING_FORCE_TO_NONPRIV_DENY to OA whitelists Unconditionally whitelisting OA registers is a security violation. Set RING_FORCE_TO_NONPRIV_DENY bit in OA nonpriv slots, so that OA registers don't get whitelisted by default after probe, gt reset, resume and engine reset. (cherry picked from commit 90511bdcfda97211c01f1d945d4ea616578d8fca)

In the Linux kernel, the following vulnerability has been resolved: drm/imagination: Count paired job fence as dependency in prepare_job() The DRM scheduler's prepare_job() callback counts the remaining non-signaled native dependencies for a job, preventing job submission until those (plus job data and fence update) can fit in the job queue's CCCB. This means checking which dependencies can be waited upon in the firmware, i.e. whether they are backed by a UFO object, i.e. whether their drm_sched_fence::parent has been assigned to a pvr_queue_fence::base fence. That happens when the job owning the fence is submitted to the firmware. Paired geometry and fragment jobs are submitted at the same time, which means the dependency between them can't be checked this way before submission. Update job_count_remaining_native_deps() to take into account the dependency between paired jobs. This fixes cases where prepare_job() underestimated the space left in an almost full fragment CCCB, wrongly unblocking run_job(), which then returned early without writing the full sequence of commands to the CCCB. The above lead to kernel warnings such as the following and potentially job timeouts (depending on waiters on the missing commands): [ 375.702979] WARNING: drivers/gpu/drm/imagination/pvr_cccb.c:178 at pvr_cccb_write_command_with_header+0x2c4/0x330 [powervr], CPU#1: kworker/u16:3/47 [ 375.703160] Modules linked in: [ 375.703571] CPU: 1 UID: 0 PID: 47 Comm: kworker/u16:3 Tainted: G W 7.0.0-rc2-g817eb6b11ad5 #40 PREEMPT [ 375.703613] Tainted: [W]=WARN [ 375.703627] Hardware name: Texas Instruments AM625 SK (DT) [ 375.703645] Workqueue: powervr-sched drm_sched_run_job_work [gpu_sched] [ 375.703741] pstate: 80000005 (Nzcv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--) [ 375.703764] pc : pvr_cccb_write_command_with_header+0x2c4/0x330 [powervr] [ 375.703847] lr : pvr_queue_submit_job_to_cccb+0x578/0xa70 [powervr] [ 375.703921] sp : ffff800084a97650 [ 375.703934] x29: ffff800084a97740 x28: 0000000000000958 x27: ffff80008565d000 [ 375.703979] x26: 0000000000000030 x25: ffff800084a97680 x24: 0000000000001000 [ 375.704017] x23: ffff800084a97820 x22: 1ffff00010952ecc x21: 0000000000000008 [ 375.704056] x20: 00000000000006a8 x19: ffff00002ff7da88 x18: 0000000000000000 [ 375.704093] x17: 0000000020020000 x16: 0000000000020000 x15: 0000000000000000 [ 375.704132] x14: 0000000000000000 x13: 0000000000000000 x12: 0000000000000000 [ 375.704168] x11: 000000000000f2f2 x10: 00000000f3000000 x9 : 00000000f3f3f3f3 [ 375.704206] x8 : 00000000f2f2f200 x7 : ffff700010952ecc x6 : 0000000000000008 [ 375.704243] x5 : 0000000000000000 x4 : 1ffff00010acba00 x3 : 0000000000000000 [ 375.704279] x2 : 0000000000000007 x1 : 0000000000000fff x0 : 000000000000002f [ 375.704317] Call trace: [ 375.704331] pvr_cccb_write_command_with_header+0x2c4/0x330 [powervr] (P) [ 375.704411] pvr_queue_submit_job_to_cccb+0x578/0xa70 [powervr] [ 375.704487] pvr_queue_run_job+0x3a4/0x990 [powervr] [ 375.704562] drm_sched_run_job_work+0x580/0xd48 [gpu_sched] [ 375.704623] process_one_work+0x520/0x1288 [ 375.704658] worker_thread+0x3f0/0xb3c [ 375.704680] kthread+0x334/0x3d8 [ 375.704706] ret_from_fork+0x10/0x20 [ 375.704736] ---[ end trace 0000000000000000 ]---

In the Linux kernel, the following vulnerability has been resolved: wifi: ath9k: hif_usb: don't dereference hif_dev after re-arming firmware request ath9k_hif_request_firmware() re-arms an asynchronous firmware load via request_firmware_nowait(), passing hif_dev as the completion context, and then still dereferences hif_dev: dev_info(&hif_dev->udev->dev, "ath9k_htc: Firmware %s requested\n", hif_dev->fw_name); The re-armed callback ath9k_hif_usb_firmware_cb() runs on the "events" workqueue and, when the firmware is missing, walks the retry chain into ath9k_hif_usb_firmware_fail() -> complete_all(&hif_dev->fw_done). That releases the wait_for_completion(&hif_dev->fw_done) in a concurrent ath9k_hif_usb_disconnect(), which then kfree()s hif_dev. The trailing dev_info() in the frame that re-armed the request can therefore read freed memory (hif_dev->udev, the first field of struct hif_device_usb): BUG: KASAN: slab-use-after-free in ath9k_hif_request_firmware Read of size 8 ... by task kworker/... ath9k_hif_request_firmware ath9k_hif_usb_firmware_cb drivers/net/wireless/ath/ath9k/hif_usb.c:1247 request_firmware_work_func Allocated by ...: ath9k_hif_usb_probe drivers/net/wireless/ath/ath9k/hif_usb.c Freed by ...: ath9k_hif_usb_disconnect -> kfree drivers/net/wireless/ath/ath9k/hif_usb.c The fw_done barrier only makes disconnect wait for the firmware chain to *terminate*; it does not protect the outer ath9k_hif_request_firmware() frame that re-armed the request and keeps touching hif_dev afterwards. Drop the post-request dev_info(): it is the only use of hif_dev after the async request is armed, and it is purely informational (the dev_err() on the failure path runs only when request_firmware_nowait() did not arm a callback, so hif_dev is still alive there). This was first reported by syzbot as a single, non-reproduced crash that was later auto-obsoleted, and was independently rediscovered by the reFuzz fuzzer, which produced a C reproducer (USB-gadget connect/disconnect of an ath9k_htc device whose firmware download fails). The vulnerable code is unchanged and still present in v7.1-rc6, where the slab-use-after-free reproduces under KASAN once the (sub-microsecond) race window is widened.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: mgmt: fix locking in unpair_device/disconnect_sync Dereferencing RCU-protected pointers outside critical sections is invalid and may lead to UAF. Take hdev->lock for hci_conn lookup and hci_abort_conn(). Don't use RCU to ensure the conn is fully initialized at this point.

In the Linux kernel, the following vulnerability has been resolved: fuse-uring: fix race between registration and connection abortion This fixes this race: - thread a: io_uring_enter -> register sqe -> fuse_uring_create_ring_ent -> allocate ent but doesn't grab queue_ref yet - thread b: fuse_conn_destroy() -> fuse_chan_abort() -> fuse_uring_abort() is a no-op due to queue ref being 0 - thread a: grabs the queue_ref, queue_ref is now 1, rest of fuse_uring_do_register() logic executes - thread b: fuse_chan_abort() returns, fuse_chan_wait_aborted() now runs and calls "wait_event(ring->stop_waitq, atomic_read(&ring->queue_refs) == 0);" The abort/unmount thread will hang indefinitely in unkillable state as nothing will decrement queue_refs or wake stop_waitq, and the ring, queue, and ent are leaked. Fix this by checking fch->connected under fch->lock after the created ent has grabbed a ref count on the queue. This ensures that in the scenario above, it is guaranteed that we either release the queue ref and wake up stop_waitq (in case fuse_chan_wait_aborted() is already waiting) in fuse_uring_do_register() when we detect !fch->connected, or if the connection is aborted after the check, it is guaranteed that the async teardown worker will be running in the background cleaning up ents and decrementing the ent's ref on the queue, which will unblock the eventual queue and ring teardown.

In the Linux kernel, the following vulnerability has been resolved: super: fix emergency thaw deadlock on frozen block devices do_thaw_all_callback() calls bdev_thaw() while holding sb->s_umount exclusively. If the block device was frozen via bdev_freeze() dropping the last block layer freeze reference calls fs_bdev_thaw() which reacquires s_umount: do_thaw_all_callback(sb) super_lock_excl(sb) # holds sb->s_umount bdev_thaw(sb->s_bdev) mutex_lock(&bdev->bd_fsfreeze_mutex) # bd_fsfreeze_count drops 1 -> 0 bd_holder_ops->thaw == fs_bdev_thaw get_bdev_super(bdev) bdev_super_lock(bdev, true) super_lock(sb, true) down_write(&sb->s_umount) # same task: deadlock The emergency thaw worker deadlocks against itself holding both s_umount and bd_fsfreeze_mutex. That fscks any subsequent unmount, freeze, or thaw of that filesystem and block device. [ 81.878470] sysrq: Show Blocked State [ 81.880140] task:kworker/0:1 state:D stack:0 pid:11 tgid:11 ppid:2 task_flags:0x4208060 flags:0x00080000 [ 81.884876] Workqueue: events do_thaw_all [ 81.886656] Call Trace: [ 81.887759] <TASK> [ 81.888763] __schedule+0x579/0x1420 [ 81.890372] schedule+0x3a/0x100 [ 81.891794] schedule_preempt_disabled+0x15/0x30 [ 81.893848] rwsem_down_write_slowpath+0x1ea/0x900 [ 81.895191] ? __pfx_do_thaw_all_callback+0x10/0x10 [ 81.896528] down_write+0xbd/0xc0 [ 81.897505] super_lock+0x91/0x180 [ 81.898457] ? __mutex_lock+0xa99/0x1140 [ 81.900748] ? __mutex_unlock_slowpath+0x1f/0x400 [ 81.902069] bdev_super_lock+0x5b/0x150 [ 81.903132] get_bdev_super+0x10/0x60 [ 81.904042] fs_bdev_thaw+0x23/0xf0 [ 81.904755] bdev_thaw+0x82/0x100 [ 81.905484] do_thaw_all_callback+0x2c/0x50 [ 81.906298] __iterate_supers+0x5d/0x130 [ 81.907067] do_thaw_all+0x20/0x40 [ 81.907739] process_one_work+0x206/0x5e0 [ 81.908545] worker_thread+0x1e2/0x3c0 [ 81.909339] ? __pfx_worker_thread+0x10/0x10 [ 81.910171] kthread+0xf4/0x130 [ 81.910799] ? __pfx_kthread+0x10/0x10 [ 81.911528] ret_from_fork+0x2e2/0x3b0 [ 81.912259] ? __pfx_kthread+0x10/0x10 [ 81.913010] ret_from_fork_asm+0x1a/0x30 [ 81.913806] </TASK> bdev_super_lock() even documents the violated requirement with lockdep_assert_not_held(&sb->s_umount). Acquiring bd_fsfreeze_mutex under s_umount also inverts the bd_fsfreeze_mutex vs. s_umount ordering established by bdev_{freeze,thaw}() and can thus ABBA against a concurrent block-layer freeze even when the recursive path isn't hit. Fix this by not holding s_umount around the bdev_thaw() loop at all. Pin the superblock with an active reference instead as filesystems_freeze_callback() does. The active reference keeps the superblock from being shut down and so ->s_bdev stays valid without holding s_umount. The block-layer-held freeze is dropped by fs_bdev_thaw() with FREEZE_MAY_NEST | FREEZE_HOLDER_USERSPACE exactly as a regular unfreeze would and thaw_super_locked() handles filesystem-level freezes as before. The emergency thaw path has deadlocked like this in one form or another for a long long time but the current exclusively-held shape dates back to commit [1] where thaw_bdev() already ended in thaw_super() with s_umount held by do_thaw_all_callback().

In the Linux kernel, the following vulnerability has been resolved: iio: core: fix uninitialized data in debugfs If *ppos is non-zero then simple_write_to_buffer() will not initialize the start of buf[]. Non zero values for *ppos aren't going to work anyway. Test for them at the start of the function and return -EINVAL.

A Linux kernel vulnerability in version 6.12 related to the scheduler's handling of core scheduling forced idle states has been resolved. The issue involved spurious warnings triggered when a runnable task dropped to a lower scheduling class without a specific enqueue flag, due to core scheduling behavior bypassing expected code paths. The fix gates these warnings to avoid false positives when core scheduling forces idle states.

CVE-2026-68200 is a use-after-free vulnerability in the Linux kernel ALSA timer subsystem. It occurs because the timer callback can be re-entered concurrently without proper serialization, leading to a race condition where an instance may be freed while its callback is still executing. This flaw affects the ALSA userspace-driven timer interface and can be triggered by any user able to open /dev/snd/timer. The issue has been fixed in Linux kernel version 6.12.

A vulnerability was determined in mifi lossless-cut up to 3.69.0. Affected by this issue is some unknown functionality of the file src/main/httpServer.ts of the component Built-in HTTP API Service. Executing a manipulation can lead to server-side request forgery. The attack requires access to the local network. This attack is characterized by high complexity. The exploitation is known to be difficult. The exploit has been publicly disclosed and may be utilized. This patch is called 260802348955231442c4bae6c2d9d8ede947af0a. It is best practice to apply a patch to resolve this issue. The project maintainer provides this view: "I'm not sure that this is a critical vulnerability, because it is behind an experimental CLI flag and the NTLM behavior isn't really a LosslessCut bug." The CVSS vector reflects the high level of pre-requisites.

PraisonAI is a multi-agent teams system. Versions prior to 0.1.4 of the PraisonAI Platform API have two authorization failures that together break workspace isolation. The service layer for issues and projects performs global primary-key lookups without checking workspace ownership, so any authenticated user can read, modify, and delete resources in any workspace just by swapping UUIDs in their API requests. On top of that, every member management endpoint (add, update role, remove) only requires `min_role="member"`, which lets any workspace member promote themselves to owner and kick out the original owner. A low-privilege member of one workspace can steal data from every other workspace and take over any workspace they belong to. Both issues come from the same gap: the route layer pulls `workspace_id` from the URL and verifies membership, but the service layer ignores the workspace scope for resource lookups and ignores the caller's role level for member operations. The `require_workspace_member()` dependency does its job correctly. The problem is that the service layer doesn't use the information it provides. Version 0.1.4 of the PraisonAI Platform API patch the issue.

Agentic AI armed attackers first, but it also put real building power in defenders’ hands. Here’s what security practitioners built in two days at Black Hat USA 2026, and how the CyberAgents Exchange keeps that work compounding long after the event. Key takeaways Building defensive cybersecurity tooling no longer requires a developer. Agentic tooling drove the cost of finding and exploiting a vulnerability down to 1990s levels; it also removed the engineering barrier that kept defenders from building the automation they’ve always wanted. The unglamorous work won the room: triage, reconciliation, toil. Given two days and a requirement to publish, practitioners at SWARM developed agents for prioritization, cross-tool reconciliation, and the unglamorous toil they recognize from their own environments. All SWARM builds live on the CyberAgents Exchange , source repos attached. Every component built at SWARM is published open source with its source repository attached, so the next team facing the same problem starts from working code instead of a blank editor. Another security team already built the AI agent you need You may not realize it, but somewhere out in the ether, there’s a security team facing the same challenge as you. The only difference is they just solved the problem with agentic AI. The problem is, you have no way to find out the solution even exists. It could be an asset inventory that three systems describe in three different ways; a findings queue nobody has the hours to work; or a “critical” that a platform upgrade quietly neutralized six months ago, still sitting there waiting for someone to prove it. Somebody has already built the thing you keep meaning to build. Now it’s time we help you find it. At Black Hat USA 2026, nearly 100 registrants had the opportunity to come together for 48 hours and solve both halves of that problem: identifying a key operational pain point and building the fix. Problems like those got solved at SWARM, and the fixes are sitting on the CyberAgents Exchange right now, open source, with their source repositories attached. One team built the agent that works out which handful of fixes retires the most risk across thousands of findings. Another correlated two scanners to tell whether a flaw in the code is even reachable in the running application. A third made the case that a finding had already been mitigated, with evidence an auditor would accept. You can download and deploy any of them today. Agentic AI doesn’t just arm attackers Black Hat’s keynote stage spent this year focused on one theme: the plummeting cost of cyber offense in the agentic AI era. The price for an attacker to find and exploit a vulnerability is at lows the industry hasn’t seen since the 1990s, when a working exploit meant weeks of expert reverse engineering. Now all it takes is an afternoon and a subscription. The artisanal exploit isn’t rare anymore. True. But neither is the defender who can build. The same agentic tooling that’s arming attackers puts real building power in everyone’s hands, and that half of the story got almost no airtime. Security automation used to require the work and ongoing maintenance of skilled engineers. Now practitioners who understand the problem can build the fix. Inside the conference room at the Mandalay Bay where Tenable hosted our inaugural SWARM event, the proof of that was on every table. The winning ranking engine ships as a skill that runs on the Python standard library alone — no packages, no install step, no build pipeline. Point it at the bundled demo estate and it answers “what should we fix first?” in seconds. That’s a deliverable a practitioner can produce and a colleague can run, and two days was enough. The attacker-defender asymmetry doesn’t stem from a lack of talent or willingness. Offensive cyber capabilities compound because the tooling circulates: it’s built once, forked, passed on, or sold to the next threat actor to leverage in their attack. Meanwhile, defenders…

OpenAI is rolling out a more reliable version of ChatGPT GPT-5.6 Sol for Plus and Pro users, while Free users are getting unlimited text chats with GPT-5.6 Luna. [...]

### Summary Four inline processors in pymdown-extensions contain regular expressions with exponential backtracking. A single untrusted Markdown line under 50 bytes drives `markdown.markdown()` into unbounded CPU on the rendering thread (seconds at ~45 bytes, growing exponentially with each added character). All four fire in the extension's **default configuration** and are reachable through the documented public API. The `caret`/`tilde`/ `betterem` blow-up was introduced by the emphasis-pattern rewrite in PR #2547 (first released in **10.13**, Dec 2024) — earlier releases used a linear `(.+?)` / `([^\s]+?)` content group — and is present through **11.0** (latest); `magiclink`'s host pattern is long-standing and affects effectively all releases. Likely **CWE-1333 (Inefficient Regular Expression Complexity)**. This is a distinct issue from CVE-2025-68142 (ReDoS in `pymdownx.blocks.caption`, `RE_FIG_NUM`, fixed in 10.16.1): different extensions, different regexes, and a different root cause (delimiter-run partition ambiguity rather than a `.`/`\.` typo). ### Details Four regexes share, or closely mirror, a vulnerable shape — an inner group that can partition a run of the delimiter character into `{2,}`-sized pieces in exponentially many ways, wrapped in a lazy `+?` that must fail before the engine can give up: | Extension | Regex | Location (`11.0`) | |---|---|---| | `pymdownx.caret` (superscript `^…^`) | `SUP2` | `pymdownx/caret.py:56` | | `pymdownx.tilde` (subscript `~…~`) | `SUB2` | `pymdownx/tilde.py:55` | | `pymdownx.betterem` (underscore `_…_`) | `SMART_UNDER_EM2` (default) | `pymdownx/betterem.py:93` | | `pymdownx.magiclink` (bare-URL autolink) | `RE_LINK` | `pymdownx/magiclink.py:56` (host at `:59`) | `pymdownx/caret.py:56` (`pymdown-extensions 11.0`): ```python SUP2 = r'(?<!\^)(\^)(?![\^\s])((?:[^\^\s]|\^{2,})+?)(?<![\^\s])(\^)(?!\^)' ``` The content group `(?:[^\^\s]|\^{2,})+?` matches a run of carets only via the `\^{2,}` branch. A run of *k* carets can be split into ≥2-length pieces in exponentially many combinations; when no caret can serve as a valid closing delimiter (the trailing `(?<![\^\s])(\^)` cannot be satisfied), the engine explores every partition before failing. `SUB2` (tilde) and `SMART_UNDER_EM2` (betterem) are the same construct for `~` and `_`. In `betterem` the default `smart_enable='underscore'` routes underscores to `SmartUnderscoreProcessor` → `SMART_UNDER_EM2` (`betterem.py:93`), which is the default-reachable, API-exploitable pattern; the non-smart `UNDER_EM2` (`:69`, used only when `smart_enable` is `asterisk`/`disable`) shares the shape but did not reproduce through the public `markdown.markdown()` pipeline on the tested payload, so a fix and regression test should target `SMART_UNDER_EM2`. `pymdownx/magiclink.py:59` has the analogous ambiguity in the host portion, where overlapping character classes let a run of dots be grouped exponentially: ```python (?:ht|f)tps?://[^_\W][-\w]*(?:\.[-\w.]+)* # host: '\.' and '[-\w.]' inside (?:...)* both match '.' ``` `SUP2`/`SUB2`/`SMART_UNDER_EM2` are applied at each delimiter occurrence via the default `PatternSequenceProcessor` subclasses (`pymdownx/util.py`); `RE_LINK` is applied by `MagiclinkPattern` (registered unconditionally at priority 85). In all four cases, rendering `markdown.markdown(src, extensions=[ext])` on untrusted `src` in default configuration is sufficient to reach the regex. ### PoC Single self-contained script; runs against the pinned release in an ephemeral env. Non-destructive — the input is ordinary Markdown text; the impact is CPU/time (a per-render alarm caps each attempt so the script terminates). ```python import signal import time from importlib.metadata import version import markdown print(f"# pymdown-extensions {version('pymdown-extensions')} / markdown {version('markdown')}") CAP = 5.0 # a single render exceeding this is treated as a hang class Timeout(Exception): pass def render(ext, text): signal.signal(signal.SIGALRM, lambda *_: (_ for _ in ()).throw(Timeout())) signal.setitimer(signal.ITIMER_REAL, CAP) t = time.perf_counter() try: markdown.markdown(text, extensions=[ext]) return time.perf_counter() - t except Timeout: return None finally: signal.setitimer(signal.ITIMER_REAL, 0) # ext -> (malicious builder, benign builder [valid & closed], ramp, hang count) CASES = { "pymdownx.caret": (lambda n: "^a" + "^" * n + "b", lambda n: "^" + "a" * n + "^", [24, 30, 36], 44), "pymdownx.tilde": (lambda n: "~a" + "~" * n + "b", lambda n: "~" + "a" * n + "~", [24, 30, 36], 44), "pymdownx.betterem": (lambda n: "_a" + "_" * n + "b", lambda n: "_" + "a" * n + "_", [24, 30, 36], 44), "pymdownx.magiclink": (lambda n: "http://a" + "." * n + " ", lambda n: "http://" + "a" * n + ".com ", [28, 32, 36], 40), } repro = [] for ext, (evil, benign, ramp, hang) in CASES.

Welcome to this week’s edition of the Threat Source newsletter. Metaphor is a powerful tool for understanding emerging issues in cybersecurity. Framing the unfamiliar in terms of the well understood helps us remove the burden of extraneous detail to draw focus to the real issues. Recent reports of offensive AI agents "escaping" their sandbox environments to attack external systems have forced the industry into a moment of rapid sense-making. How we interpret this event doesn’t just reflect our perspective, but shapes our long-term response. We can imagine three different narratives for interpreting the escape of autonomous agents. The innovation narrative: We can marvel at the advance of technology, considering these agents as plucky entities with a thirst for knowledge and resources, who found clever ways to sneak out of their digital confines. The response: If the AI is a naughty child, our reaction is one of mild disapproval or gentle rebuke where better “parenting” (guardrails) is appropriate. It minimizes the threat, framing it as the unexpected hijinks of a brilliant new technology. The safety narrative: Imagine a breeder who has trained the world's most intelligent guard dogs. Despite high fences and barriers, their ability to identify weaknesses allows them to escape, run riot and menace local businesses. The response: The framing shifts to biology and inherent danger. We question if the breeder can be trusted and whether such inherently wild technology requires strict regulation to ensure public safety. The liability narrative. Finally, we can view the incident as an industrial accident. A company developing a new chemical substance experiences a containment failure. The agent leaks into the environment through an unforeseen mechanism causing damaging pollution to those in its path. The response: The framing invokes the language of the lawyer, implying negligence, lack of duty of care, and financial liability for the harm caused. The conversation moves from innovation to corporate responsibility, regulatory oversight, and the diligent management of hazardous materials. First impressions matter. Sensemaking shapes how we perceive incidents. Our initial perceptions of an incident dictates how we react to similar situations in the future. If we consider that the escape of an AI agent is an example of innovative autonomous thinking, then we will continue to prioritise speed over safety. Conversely, if we consider the issue as one of failed hazard containment, then we shall build a future of enforced safety standards backed by legal liability. There is no right or wrong metaphor. Our interpretation depends on our personal system of beliefs. Personally, I would argue that the unintentional release of something that causes damage is, at its core, a failure of engineering and foresight. Words shape our reactions. Metaphors help us understand new situations and tap into our prior experience to address problems that have yet to fully manifest. We need cognitive tools to help our understanding, but we must be aware of the metaphors that are being foisted upon us which may shape our thinking. Excuses and the trivialisation of incidents may hide failings, allowing them to accumulate until they manifest as more damaging incidents. Conversely, overreacting risks stifling research and diverting resources away from more relevant and pressing threats. New threats require new ideas. Metaphor helps us make sense of a changing world, but in this new era, the person who shapes the narrative controls the strategy. The one big thing Cisco Talos released a data-driven analysis of how adversaries are weaponizing AI in the wild. By analyzing prompt logs left behind on endpoints, we found threat actors successfully bypassing guardrails to use AI as malicious software engineers, criminal force multipliers, and vulnerability research accelerators. While novice hackers use AI to cobble together buggy malware, sophisticated actors are building highl…

A 26-year-old Canadian man once described as one of the most consequential cybercrime threat actors of 2024 has pleaded guilty to computer fraud and conspiracy to hack and extort more than 165 organizations that used the cloud provider Snowflake . Connor Riley Moucka , of Kitchener, Ontario, also admitted to stealing call and text history records of more than 100 million AT&T customers. A surveillance photo of Connor Riley Moucka, a.k.a. “Judische” and “Waifu,” dated Oct 21, 2024, 9 days before Moucka’s arrest. This image was included in an affidavit filed by an investigator with the Royal Canadian Mounted Police (RCMP). The U.S. Justice Department said between February and October 2024, Moucka and co-conspirators used stolen login credentials to steal cloud-hosted data belonging to at least 165 customers of a U.S.-based software-as-a-service company. The hackers targeted stolen credentials for Snowflake customer accounts that did not enforce multi-factor authentication, and extorted or attempted to extort a host of well-known companies, including TicketMaster, Lending Tree, Advance Auto Parts and Neiman Marcus. Snowflake responded to the data thefts by increasing password complexity requirements and enforcing multi-factor authentication. Moucka adopted new nicknames frequently — sometimes operating multiple identities concurrently — but two of his best-known monikers were “ Judische ” and “ Waifu .” Judische’s admitted role in the Snowflake data thefts was first documented by KrebsOnSecurity in a September 2024 story about the overlap between Western, English-speaking cybercriminals and extremist groups that harass and extort minors into harming themselves or others. That September 2024 story identified Judische as a software engineer from Ontario who has been involved in numerous data breaches and voice phishing attacks against U.S. companies since at least 2020. A little more than a month later, Canadian authorities arrested Moucka on a provisional warrant from the United States. The government says Moucka and others used their unauthorized access to steal billions of sensitive customer records and download terabytes of information, “including individuals’ non-content call and text history records, banking and other financial information, payroll records, Drug Enforcement Administration (DEA) registration numbers, driver’s license numbers, passport numbers, social security numbers and other personally identifiable information. They then extorted victims by threatening to publish data online.” Moucka also threatened and harassed government officials and security researchers who were helping to track him down. The Justice Department said the conspirators made over $2.5 million in ransom payments, and that in at least one instance, Moucka re-extorted a victim with threats of further disclosure of the victim’s stolen data. “Moucka used the stolen data of a government officer and members of a then-former government officer’s immediate family in this re-extortion attempt,” reads a statement from the Justice Department. One of Moucka’s admitted co-conspirators is Cameron “Kiberphant0m” Wagenius , a U.S. Army soldier who pleaded guilty in July 2025 to extorting AT&T and Verizon for their customer account data. Less than a month before Wagenius’s arrest, KrebsOnSecurity published a deep dive into Kiberphant0m’s various Telegram and Discord identities over the years, revealing how the owner of the accounts told others they were in the Army and stationed in South Korea. One of several selfies on the Facebook page of Cameron Wagenius. Kiberphant0m also re-extorted victims. Immediately following Moucka’s arrest, Kiberphant0m posted on hacker forums what he claimed were the AT&T call logs for then President-elect Donald Trump and for then Vice President Kamala Harris, as well schematics allegedly stolen from the U.S. National Security Agency (NSA). Wagenius is set to be sentenced on September 3, 2026. The government says he faces a maxim…

A flaw has been found in Insta InstaKNXServiceApp 1.2.3.1469. Affected by this issue is the function CreateWebClientAndDownloadFileList of the component Firmware Update Handler. Executing a manipulation can lead to insufficient verification of data authenticity. It is possible to launch the attack remotely. Attacks of this nature are highly complex. The exploitation is known to be difficult. There are still doubts about whether this vulnerability truly exists. The vendor explains: "[T]he entry suggests that the device does not verify the authenticity of the update. This is incorrect. The device performs an authenticity check after the complete update has been transferred. If this check fails, the update is not installed and the previously installed firmware continues to run".

For security researchers studying unconventional side-channel attacks, acoustic keylogging is something of a Hello World: a foundational problem that’s been tackled many times. A recent paper authored by researchers across three Japanese universities cites six previous studies on the topic that date as far back as 2004. While earlier experiments showed theoretical promise, they came with real-world caveats so severe that it made them all but impractical for actual espionage. The authors of this latest study, however, claim to have overcome most of those limitations. Today, we look at how they pulled it off, and assess whether their method holds up in real-world scenarios. What makes this new approach different? Previous acoustic keylogging techniques were fundamentally flawed. Best-case scenarios required prior training on the target’s specific keyboard model. Worst-case scenarios required a complex microphone array to isolate the subtle acoustic differences between keystrokes. Crucially, almost all prior models failed outside silent environments, which rendered the attack vector virtually useless. The Japanese research team demonstrated reliable keystroke interception even if the target was sitting nearby in a public space, sound was being recorded in an online meeting, or the researchers were using a contact microphone to eavesdrop through a wall. All this with strong model accuracy and a minimal training dataset. Their process needs a sample of just 150 to 200 keystrokes to reach a 99% accuracy rate for subsequent typing. Attack scenarios and core methodology proposed by the Japanese researchers. Source How to crack 200 keystrokes in under 50 iterations To understand how the researchers achieved such high accuracy and adaptability, we have to look at their audio processing pipeline. Their analysis begins by automatically segmenting a raw recording into discrete keystrokes. This data is then passed through a specialized algorithm that simplifies the subsequent audio analysis. Next, the system clusters together acoustically similar signals. The assumption is that the members of one cluster map to the exact same key. One particularly intriguing takeaway was isolating the spacebar sound from all the rest. Because the spacebar produces a distinctly unique sound profile compared to other keys, identifying it provides reliable word boundaries. This streamlines the next phase: feeding the preprocessed acoustic data into specialized language models for inference. Yes, the method relies on not one but two language models. The first model performs multiple passes over the audio stream to map acoustic signatures to potential keyboard characters. During each pass, the model leverages dictionaries to hypothesize character mapping, and check whether the resulting text aligns with standard words. The second model handles the final refinement pass: it ingests thoroughly pre-processed data rather than raw inputs. The method doesn’t stop there: unrecognized keystrokes undergo manual analysis, with analysts injecting educated guesses before re-running the recognition pipeline once again. The goal of looping through these multiple iterations is to achieve complete recognition across the keyboard from an ultra-compact dataset of ideally no more than 200 captured keystrokes. This marks a major shift from legacy methods, which relied on massive training datasets. Clustering the sounds of keystrokes permits grouping similar acoustic profiles together prior to recognition. Notice how distinctly the spacebar sounds stand out: they make subsequent text reconstruction vastly simpler. Source Research results To validate their theoretical model, the researchers created an experimental testing setup: Clustering the sounds of keystrokes permits grouping similar acoustic profiles together prior to recognition. Notice how distinctly the spacebar sounds stand out: they make subsequent text reconstruction vastly simpler. Source The team tested four distinct la…

(Video) In this podcast, we share insights from Edna Conway, a recognized leader in cybersecurity and supply chain resilience with over 40 years of experience in the field. The post Podcast: Compliance Won’t Save You: The Future of Cyber Risk with Edna Conway appeared first on SecurityWeek .