Cyber Digest
A daily roundup of key cybersecurity developments across sectors
Executive Summary
The weekend's most consequential story is criminal-on-criminal, and it threatens to turn one gang's payment records into the next round of extortion. ShinyHunters has taken over the dark-web leak site of the Cl0p ransomware operation and is using Cl0p's own pressurising machinery against it: the defaced page carries an eight-figure demand the attackers describe as "2.333%" of Cl0p's net worth, a threat to raise that figure every 24 hours Cl0p fails to respond, and — by Monday — a requirement for a public apology. The material risk sits with Cl0p's *victims* rather than Cl0p: ShinyHunters says it will publish records showing which companies paid, how much they paid and which Bitcoin addresses were used, which would hand every one of those organisations a second extortion attempt built on the first one's evidence. The second thread is supply-chain provenance, and it failed exactly where the ecosystem now places its trust. CloudSEK's account of the GHAPPIER loader shows an attacker using a legitimate maintainer account and GitHub Actions OIDC trusted publishing to ship malware from a repository CI that npm was designed to believe: the Sigstore attestation is genuine and still sitting in the public log, naming the attacker's commit, because provenance proves where an artefact was built and not whether its source was honest. Third, the Windows endpoint story turned on signed components, with a fake LastPass Authenticator installer reaching SYSTEM through a Microsoft-signed kernel driver that terminates 145 security processes, and a new PowerShell backdoor, TASK#STOMP, harvesting documents, Wi-Fi credentials and clipboard contents through redundant, mutually-watchdogged C2 channels.
The Australian signal in the window is a quiet one from the ACSC and a pointed one from the ASD. The ACSC has published no new alert or advisory since its 18 September WaterPlum advisory — the newest items on the site remain the 17 September network segmentation and segregation publications (an overview plus an anti-patterns companion) and the 15 September Active Directory detection guidance — so today's DPRK material runs through the ASD's own messaging rather than a fresh national product. On 22 September the ASD joined Japan, Germany and the United States in warning that North Korean actors are posing as recruiters to target IT professionals, but declined to say whether any Australian individual or organisation has been caught up in WaterPlum, telling iTnews only that the tactics "are being used against organisations across the world" and that advisories are published when doing so is assessed to be in Australia's interests. The figures behind that refusal are substantial: 30,000-plus infected devices in more than 100 countries, 7,000-plus drained cryptocurrency wallets and 1.7 billion yen (about US$10.71 million) transferred to the DPRK. Australia also sits on the Multilateral Sanctions Monitoring Team, whose latest report estimates DPRK overseas labour earned the regime US$450 million to US$800 million in 2025 across at least 17 countries, with 80 to 90 per cent of wages confiscated — the same employment pipeline, read from the other end. For Australian defenders the operative obligation has not changed: Critical infrastructure operators should read the tanker boardings and the Colorado water manipulation as the SOCI Act risk picture rather than as foreign news.
Over seven days the supply-chain story has migrated from the install script to the build pipeline. Four campaigns now line up in sequence: TanStack in May, Brevo's injection of ClickFix scripts across customer sites, PhantomRaven's LLM-written stealer, `indexed-btree` hiding its loader inside ordinary runtime code, and now GHAPPIER abusing OIDC trusted publishing so that a valid attestation ships with malicious source. Each step moves the compromise one layer deeper into something the ecosystem treats as a trust anchor — dependency, then maintainer account, then CI identity — and each step leaves the previous defence intact and irrelevant. The lesson defenders should draw is that signature and provenance checks answer a narrower question than they are asked to answer. Extortion is meanwhile fragmenting at both ends and, for the first time in this run, feeding on itself. N0n posted a 108 GB claim against Fanatics last week with a data-broker profile; Settra runs a smaller hand-operated double-extortion play; and ShinyHunters has spent the weekend attacking a rival gang's infrastructure, threatening to publish Cl0p's payment ledger. That last development is the one to watch: payment records are the difference between a victim knowing it settled and the whole market knowing what it settled for. Operationally, the tempo of critical-infrastructure probing is the other through-line. Colorado's governor's office confirmed foreign actors changed equipment settings at two small water utilities in late August, disabling alarms and remote access and altering pumping cycles; the Coast Guard boarded two oil tankers heading for Texas in the same month after indications their networks were compromised; and the Port of Los Angeles reported blocking more than 120 million attacks in August. None of these produced a service failure, and that is precisely what makes them worth logging — the access was achieved, the disruption was not attempted. For the week ahead: whether Cl0p's payment ledger actually appears, whether GHAPPIER spawns a second victim class as lockfiles pinned to version 0.2.21 surface in CI logs, and whether the DPRK recruiter campaign produces a named Australian victim now that the ASD has publicly declined to rule one out.
Incident Map
Global (Macro) 4 stories
ShinyHunters Hijacks the Cl0p Leak Site and Threatens to Publish Which Victims Paid
The ShinyHunters extortion group took over the dark-web leak site of the Cl0p ransomware gang over the weekend and is running Cl0p's own pressure tactics against it. The defaced page carries an extortion demand described as "2.333%" of Cl0p's net worth — an unspecified eight-figure sum — and states that the figure will rise every 24 hours that Cl0p fails to respond; by Monday the terms had expanded to require a public apology. A message posted on Sunday named three people identified as Cl0p operators, all previously named in public reporting, and demanded the proceeds of Cl0p's recent Oracle E-Business Suite campaign "plus more". The demand that matters most to defenders is the threat to publish records showing which companies paid Cl0p, how much they paid and which Bitcoin addresses were used: any organisation that settled with Cl0p during the E-Business Suite campaign would be exposed to a second extortion attempt armed with proof of the first. ShinyHunters says the feud began with Cl0p's unauthorised use of a vulnerability ShinyHunters had published proof-of-concept code for, and with threats against one of its members. Cl0p replied on Monday asking ShinyHunters to make contact.
GHAPPIER Supply-Chain Campaign Abuses npm Trusted Publishing to Ship a Loader With Valid Provenance
Attackers used the maintainer account of the npm package `@dforge-core/dforge-mcp` for 105 minutes on 9 September to publish a previously unreported loader that CloudSEK tracks as GHAPPIER, and the release carried a genuine build attestation. Version 0.2.20 broke installation; 0.2.21 shipped the loader and held as the latest version for 35 minutes and 38 seconds. The attacker could already push to the main branch and changed three lines so that any push triggered the release workflow, then rewrote that workflow 14 minutes later so it could publish unattended. The build ran through GitHub Actions with OIDC trusted publishing, so the attestation sits in Sigstore's public log naming the attacker's own commit, and the release would pass `npm audit signatures`. CloudSEK's framing is the one to keep: "provenance attests where an artefact was built, not whether its source was honest." The loader was a single line inside a 99 KB file, opening a four-stage chain that ended in a general-purpose remote shell which deleted itself as it ran; it fired when the MCP server was launched rather than on install, so systems that installed 0.2.21 without starting it did not execute it. The activity spans at least 65 public repositories, 73 infected files and 22 accounts, and a second payload in another victim's repository matched PolinRider, the campaign OpenSourceMalware has tracked since March 2026 and which other researchers attribute to North Korea; a configuration step reads from an empty Ethereum transaction costing about $0.20, leaving no domain to seize. CloudSEK advises pinning 0.2.22, treating a lockfile that pins 0.2.21 as an indicator in itself, and alerting on changes to a release workflow's trigger block. No OSV or GitHub advisory had appeared at the time of the report.
Fake LastPass Authenticator Installer Reaches SYSTEM With a Microsoft-Signed Kernel Driver
A fake LastPass Authenticator installer hosted on GitHub installs a Windows kernel driver that disables antivirus and endpoint security before a stealer runs, according to research from LastPass and Delphos Labs published on 17 September. The lure is a counterfeit GitHub page (`github.com/LastPass-Authenticator`) that ranks in search results for download terms and imitates a genuine product page; the real application is distributed from `lastpass.com` and the official app stores. Clicking through leads to an attacker server serving a padded ZIP — the samples seen were 148 MB and 127.9 MB, inflated with junk files so size-limited scanners skip them — containing a renamed copy of Microsoft's `vsdbg.exe` beside a malicious `vsdbg.dll`. Windows loads the attacker's DLL from the same folder by DLL side-loading, after which the loader makes three attempts at administrator rights, reaches SYSTEM and installs the driver as a service. The driver, named `Alinubx.sys`, carries a list of 145 antivirus and security process names and terminates each one from kernel level, below where those tools run. It is signed through Microsoft's Windows Hardware Compatibility Publisher chain with a March 2023 signing date, scored zero detections on VirusTotal when checked in August and was absent from Microsoft's blocked-driver list — a bring your own vulnerable driver (BYOVD) technique. With defences down, the stealer harvested saved passwords from more than two dozen browsers, cryptocurrency wallet files, Discord, Steam and Telegram sessions, Windows Credential Manager contents and files named like "password", "seed" or "recovery". LastPass says none of its systems, services or customer vaults were affected.
TASK#STOMP Backdoor Harvests Documents, Wi-Fi Credentials and Clipboard Through Twin Watchdogged C2 Channels
Securonix researchers have detailed a campaign they call TASK#STOMP that delivers a PowerShell backdoor capable of "automatically harvesting and exfiltrating business documents, watching the filesystem for new files in real time", stealing Wi-Fi passwords and clipboard contents, capturing screenshots and accepting arbitrary remote commands through two redundant, token-authenticated command-and-control servers. The chain begins with `wscript.exe` executing an encoded VBScript staged on the victim's desktop under a random filename, `95c9050t66.vbs`; the initial access path is unclear and may have been email-based phishing. The script establishes persistence through scheduled tasks deliberately named Local Credential Manager, Network Audio Service, Windows Display Manager and Device Credential Handler so they blend into normal operating-system activity, and adds a backup persistence route through the Startup folder script `msdiag.vbs`. Two further modules run as separate processes — `sys_loader.ps1`, decoding `diag_pack.dat`, and `win_conn.ps1`, decoding `win_conn_cfg.dat` — communicating with infrastructure at `corecloudfileshare[.]xyz` or `attachmentsharingdrive[.]xyz`. The pair hold a mutual-watchdog relationship, each restarting the other if it stops, so terminating one branch does not end the intrusion. Timestomping and cleanup routines are present to frustrate casual administrative review and forensic analysis.
Defence 1 story
Jade Sleet Tied to an Indian IT Services Breach Using macOS Backdoors From the KelpDAO Case
SentinelOne has attributed the compromise of an India-based IT services organisation — described as a much smaller firm — to the North Korean actor Jade Sleet, tracked elsewhere as PUKCHONG, Slow Pisces, TraderTraitor and UNC4899, using macOS backdoors called FLATROOF (also known as Gaslight) and ROOFDECK that were previously seen in the March–April 2026 attack on KelpDAO's LayerZero bridge. The campaign runs through job-interview lures aimed at job seekers at the targeted company and its vendors, with targeted individuals working in DevOps, cryptocurrency or financial technology; the actor publishes GitHub repositories dressed as infrastructure-engineering projects belonging to the company it is impersonating, among them `gtn-candidate-repo`, `Northwind-IAC`, `novacart-interview` and `terraform-candidate-repo`. Those repositories contain a weaponised Terraform dependency lock file (`.terraform.lock.hcl`) pointing at attacker-controlled domains such as `registry.hashicorp-aws[.]com`, so the malware downloads when an unsuspecting developer runs `terraform init`. Both payloads are Rust-based and target ARM macOS systems: FLATROOF uses Telegram for command and control and can execute commands, move files and exfiltrate data via a Python module that collects browser data from Chrome, Brave, Firefox and Safari, Terminal histories, installed applications, system profiles, running processes and a copy of `login.keychain-db`; ROOFDECK uses the Nostr protocol for decentralised command and control, supports remote shells, file manipulation and lateral movement, persists through Launch Agents, and verifies command integrity against an embedded public key before execution. Jade Sleet was previously linked to the roughly US$1.5 billion Bybit cold-wallet theft in early 2025.
Government 2 stories
CISA Adds a Zyxel Switch Buffer Overflow to the Known Exploited Vulnerabilities Catalog
CISA added one vulnerability to its Known Exploited Vulnerabilities (KEV) Catalog on 21 September on evidence of active exploitation: CVE-2026-7273, a stack-based buffer overflow in Zyxel GS1900 series switches. The affected product line is a widely deployed small-business and branch switch family, which is why the entry matters beyond the federal estate it is formally addressed to — edge switches sit outside the patching cycles that cover servers and endpoints, and an exploited overflow on one is both a foothold and a place to persist where few monitoring tools look. The addition lands under Binding Operational Directive 26-04, CISA's risk-based vulnerability management directive, which directs federal civilian agencies to prioritise remediation of KEV entries on publicly exposed assets that grant total control after exploitation, and which sets an expectation that agencies check for pre-patch compromise rather than simply patching. CISA notes the directive binds only federal civilian agencies but encourages all organisations to adopt the same prioritisation, and continues to accept nominations for catalog entries where a CVE ID, exploitation evidence and mitigation guidance exist. The week's earlier additions — the three Linux kernel flaws added on 18 September with a 21 September federal deadline — were covered in the 19 and 20 September editions.
A Third Chinese Actor Used Spoofed News and Think-Tank Sites to Deliver Unpatched Chrome and Windows Zero-Days
Volexity has documented a third Chinese threat actor, tracked as UTA0565, using the same chained zero-days it reported on 9 September from two other Chinese APTs — CVE-2026-85046 and CVE-2026-87491 in Google Chrome alongside CVE-2026-85880 in Windows — but with a different delivery technique. In campaigns run on 3–4 September, while the vulnerabilities were still unpatched, UTA0565 sent phishing emails linking to spoofed copies of legitimate websites rather than to attacker-branded infrastructure. One Chinese-language message to Asian government entities urged recipients to publicly support the imprisoned Hong Kong activist Chow Hang-tung and amplify her voice against Chinese Communist Party suppression of a June 4 commemoration; a second masqueraded as the Center for American Progress. The spoofed domains were `chinadigitaltimes[.]top`, imitating chinadigitaltimes.net, and `americanprgoress[.]top`, imitating americanprogress.org — the typo in the second is the attacker's, not this report's. At the time of analysis the fake China Digital Times site was no longer reachable, but a Censys search showed hosting IP 96.9.125[.]52 had served a copy designed to look identical to the legitimate site. The pattern is the notable part: the phishing page and the exploit chain are one operation, so a victim who trusts the domain is delivered straight into a zero-day.
Legal Services 2 stories
Ireland's Data Protection Commission Fines Google €403 Million Over Location Data Processing
Ireland's Data Protection Commission fined Google €403 million (about US$462 million) on 21 September, closing an inquiry opened in February 2020 into how the company processed location data, and ordered the company to bring its processing into compliance within six months. The investigation examined three services and features — web and app activity, location history and location accuracy — across the period from May 2018, when the GDPR took effect, to February 2020, and considered whether the processing was lawful and fair and whether Google met the regulation's transparency and accountability requirements, including its retention of location data in those features. This is the first time the DPC, which acts as Google's lead supervisory authority because the company's European headquarters are in Dublin, has fined the platform, though the same regulator has levied repeated multi-hundred-million-euro penalties on Meta and TikTok. Deputy Commissioner Graham Doyle framed the finding around the sensitivity of the data class: location data can disclose where a person is and "a significant amount of information about an individual, including information that is inherently private". It is not established from the decision whether Google still processes any of the practices examined, and the company did not respond to a request for comment.
LinkedIn Wins a Consent Judgment Forcing Two Scraping Firms to Delete What They Took
A California federal judge finalised an agreement on Thursday between LinkedIn and two software companies, ProAPIs and joint business operator Netswift, requiring the firms to stop scraping member data at scale, stop selling and transferring it, cease accessing LinkedIn through fake accounts and delete the data already scraped. LinkedIn sued the companies and their chief executive in October 2025, alleging they built a network of bogus accounts — numbering in the millions — that scraped member, company and school information along with reactions, comments and posts, and that they created hundreds or thousands of new accounts daily, making the activity impossible to stop even though individual accounts were blocked within hours. Sarah Wright, who oversees litigation and enforcement for LinkedIn, described the outcome as vindication of the position that "your profile is yours" and that what members share is intended for their professional community rather than for an outside company to take and use. ProAPIs, which describes itself as a data pipeline platform, said it does not offer tools to scrape LinkedIn and has agreed never to do so. The judgment is the second scraping win LinkedIn has publicised this year, following action against a firm distributing a scraping browser extension five months ago.
Healthcare 1 story
Translation Vendor's Breach Reaches UnitedHealthcare Plan Members' Claims and Billing Data
United Language Group, a Minneapolis-based translation, localisation and interpretation provider, has notified the HHS Office for Civil Rights of a breach affecting the protected health information of 4,649 individuals, with the data relating to work carried out for clients UnitedHealthcare and UnitedHealthcare Global. Suspicious activity was identified in parts of its network on 9 July 2025, and the forensic investigation confirmed unauthorised access between 8 and 9 July 2025. According to notification letters, the affected data included claims, billing and other member and provider communications, and spanned names and contact information, health insurance information, health information such as diagnoses, treatment and prescriptions, Social Security numbers, financial account information including credit and debit card numbers, and driver's licence, passport, military ID and residence-permit details, with the exact combination varying by individual. The case illustrates the exposure a health plan inherits from an ancillary vendor: an interpretation service holds clinical detail and identifiers well beyond the booking data it is usually assumed to process, and the intrusion ran for two days more than a year before affected people were told. The same HHS notification round also covered breaches at Desert Pulmonary & Sleep Consultants in Arizona and Azle Cube Smiles in Texas.
Education 1 story
LMU Munich Says Enrolment Data Was Retrieved in a Breach of More Than 52,000 Students' Records
Germany's Ludwig Maximilian University of Munich is investigating a cyberattack in which an unidentified attacker accessed a system holding student enrolment data, including health insurance numbers, financial aid identifiers and bank details, and the university says it must assume the data was in fact retrieved. LMU, one of Germany's largest universities with more than 52,000 students, detected the incident on Wednesday and disclosed it on Saturday; it has not established when the access began or how long it lasted, and has not said how many people are affected or how much data was taken. The exposed records include names, dates of birth and other identifying information, contact details and LMU email addresses, bank account information, course of study and previous educational qualifications, plus numbers tied to Germany's student financial aid programme and, in some cases, information about the reasons students requested leaves of absence. Examination records, specific course content and individual academic performance were not affected. The university disconnected the affected server, brought in external cybersecurity specialists, engaged law enforcement and shut down further systems as a precaution, leaving some internal services temporarily unavailable. No attacker has been identified and the university has not said whether a ransom demand was received.
Financial Services 1 story
Smishing Wave Targets Revolut Customers Two Days After the Firm Acknowledged Its Breach
Attackers have moved to exploit Revolut's acknowledged data breach with a phishing wave aimed at its customers, according to Malwarebytes, which documented several smishing messages received by Revolut users. One message arrived on 14 September, two days after the firm confirmed the incident, and in at least one case the scam text appeared inside the same conversation thread as genuine Revolut messages on the victim's device, which is the detail that makes this campaign more effective than a generic SMS phish: no unfamiliar sender appears, and the message inherits the credibility of the thread it lands in. The text urged the recipient to follow a link to confirm their identity or face restrictions on access to their account, and a separate customer reported that opening the link led to a page soliciting further information. The pattern is the standard post-disclosure playbook — breach reporting supplies both the target list and the pretext, and the window immediately after disclosure is when recipients are most likely to respond to an identity-verification prompt — and it applies to any financial institution that publishes an incident notice. Customers of a firm that has disclosed a breach should treat identity-verification requests as suspect regardless of which thread they arrive in.
Energy & Utilities 1 story
Colorado Says Foreign Hackers Changed Settings at Two Small Water Utilities in Late August
Foreign hackers targeted and manipulated equipment at two privately owned Colorado water systems last month, according to a spokesperson for Governor Jared Polis. The utilities each serve fewer than 200 people, and the state says there was no impact to public safety or water services. The intruders changed equipment settings — disabling remote access and alarms and altering pumping cycles — and the incidents were brief, with the providers restoring control themselves and then alerting the state. The governor's office said it cannot confirm which foreign actors were involved but is aware of continuing nationwide efforts, attributed by others to an Iranian-backed group, to access drinking water and wastewater systems. The timing places the incidents after a wave of attacks on equipment at roughly 100 water entities across about a dozen states, addressed in a 21 August CISA advisory, and after an FBI statement on 30 July describing degraded water operations in earlier intrusions, including changed IP addresses and activated passwords that produced a loss of view and, in some cases, of function over connected equipment. CISA and the FBI declined to comment on the Colorado cases.
Transport 1 story
FBI and Coast Guard Boarded Two Tankers in the Gulf of Mexico After Indications Their Networks Were Compromised
US authorities are investigating whether foreign cyber adversaries were behind activity targeting ships in US waters, after the Coast Guard, working with the FBI and other federal cyber specialists, boarded a foreign-flagged commercial ship sailing toward the United States on 21 August. A Coast Guard spokesperson said the measures were designed to ensure the integrity of the vessel's operational and information technology systems "following indications that the vessel's networks were compromised by foreign cyber actors", and FBI officials confirmed a second, similar boarding on 24 August. Both vessels were oil tankers in the Gulf of Mexico en route to Texas. Officials reported no indications of operational disruption, vessel instability, danger to crew or environmental impact, and described the ship's captain, crew and shore-side corporate staff as critical partners in mitigating the threat, with the Coast Guard managing communication with port operators, owners and maritime stakeholders to keep port operations running. The boardings sit inside a wider tightening of maritime requirements, including mandatory reporting for the sector, and a specific concern flagged by analysts that foreign vessels may not meet the minimum cybersecurity standards that would prevent or mitigate such an attack.
Retail & Entertainment & Sport 3 stories
BigCommerce Warns Merchants After Third-Party App Credentials Were Used to Inject Scripts Into Stores
Ecommerce platform BigCommerce has alerted multiple merchants to data breaches after attackers compromised credentials for third-party Ribon applications and used them to inject malicious scripts into online stores. The company confirmed the credential compromise on 17 September and removed the applications immediately, and says the attacker used the compromised access to reach shopper data between 13 and 17 September. UK spirits retailer Master of Malt, one of the customers notified, told shoppers that the attacker compromised a BigCommerce application key held by Ribon and used it to reach customer data held on the platform's systems; the impacted details include full names, email addresses, phone numbers and shipping postal addresses. The case is the standard shape of app-marketplace risk: a merchant extends its storefront with a third-party integration, the integration's API key carries the merchant's scope, and a compromise at the vendor propagates straight into the merchant's customer records. Removing the apps stops the injection but does not undo the collection window, so affected merchants are in the same position as any breach victim — notification, and the expectation of follow-on phishing against the addresses and phone numbers that were taken. BigCommerce has not published a total figure for affected shoppers, and the number of merchants involved has not been disclosed.
Belgian Table Tennis and Gymnastics Federations Carry an Extortion Claim Their Own Checks Do Not Confirm
Belgium's national table tennis federation and its French-speaking branch are investigating a cyberattack after a hacker claimed to have stolen data on tens of thousands of members and users. Jean-Michel Mureau, president of the AFTT, confirmed that "there was indeed an attack" and said the Royal Belgian Table Tennis Federation was also affected, with the federation alerted on Thursday and its IT provider asked to establish what data was compromised. A statement issued on Monday said the AFTT was treating the matter seriously and had checked its own systems and those of the service provider responsible for its website, finding at that stage no evidence that its infrastructure had been compromised, and no evidence that members' phone numbers or email addresses had been taken. The attacker, using the handle Venus1337, claims to have obtained nearly 1.7 GB of data covering more than 66,800 members and 17,400 users plus club administrative accounts, and samples reportedly posted online include names, membership numbers, member and user information and records of sporting sanctions. The incident was first publicised by FrenchBreaches, a leak-tracking site. The gap between the claim and the federation's findings is the story: for a membership body, the sensitive element is not the contact list but sanction records attached to named individuals.
OpenSourceMalware Verified a Magento Payment Skimmer Hidden Behind a Benign npm Entry Point
OpenSourceMalware verified a new npm package, `homestack-cheer`, as a Magento and Stripe payment-card skimmer using deliberate dual-entrypoint smuggling. The package's `main` UMD entry runs a benign hello-world greet library, while its ESM entrypoints — `src/index.js` and `src/env_load.js` — each execute `new Function(atob('<~180KB base64>'))`, decoding multi-layer obfuscated payloads at runtime. The decoded function, `qwdsc()`, harvests card number, expiry and CVV, billing address, email address and user agent into a JSON object for exfiltration. The record is rated critical and carries OSV advisory MAL-2026-16333; OSM first recorded it on 21 September and records the affected version as all, so the version check is the consumer's to make against their own dependency graph. The technique is worth noting separately from the package: an entry point that a reviewer or a scanner inspects will show clean code, and the payload only exists after Base64 decoding at runtime, which defeats the static review most teams rely on for small utility dependencies. OSM's watch added it as one of 68 newly archived records in the 22 September harvest, alongside a cluster of Tailwind-impersonating and `chai`-named packages already tied to DPRK-nexus campaigns by the same pipeline's tagging.
AI & Frontier Technology 1 story
Of 225 Bugs Found by AI-Assisted Research, One Has Been Exploited in the Wild
Fewer than 0.5 per cent of the vulnerabilities linked to Anthropic or Project Glasswing are being exploited in the wild, according to VulnCheck researcher Patrick Garrity, who tracked 225 flaws found by Glasswing and found just one with confirmed exploitation. The finding cuts against the working assumption that AI-assisted bug discovery will translate directly into a larger exploited-vulnerability base — the CVEs are real and the volume is climbing, but the population that attackers bother with is a rounding error against it, and the likely reason is that the vulnerabilities AI tooling surfaces skew toward the complex and the hard-to-reach rather than toward the exposed, trivially triggerable classes that opportunistic actors actually use. It is a useful counterweight to the week's other AI-security results, which concern agent sandboxing and credential handling rather than newly disclosed flaws, and it argues that defenders should prioritise on exploitability evidence rather than on the provenance of the find. The caveat is the observation window: a near-zero exploitation rate over a short period says something about attacker behaviour so far, not about the eventual risk in a codebase, and the ratio will move if a single high-value flaw in the set acquires a public proof of concept.
Analytics
Source Reliability Index
| Tier | Label | Description |
|---|---|---|
| ● Tier 1 | Very High | Official / first-party |
| ● Tier 2 | High | Established cyber journalism |
| ● Tier 3 | Moderate | General tech/news media |
| ● Tier 4 | Low | Social / unverified |