Chainalysis: every user's tokens, please reset

Chainalysis breach scope balloons to all users' hashes and live tokens

Chainalysis first disclosed this incident on 30 October and said there was no evidence credentials or session tokens had been touched. On Wednesday night that changed considerably: according to The Record's report of the company's latest update, names, usernames, password hashes and live tokens for all Chainalysis users were exposed to a malicious actor. Some customers' Know Your Transaction accounts appear to have been accessed, while the company says a separate on-path password-interception method was contained and remediated. The KYT interface and API, address-screening products and the login page were taken offline; Reactor and Data Solutions had already been down since Monday, and KYT is meant to come back via customer IP allowlists. Chainalysis still says there's no evidence investigations data was accessed or attribution data manipulated. The perpetrator is unidentified, and the company's line about an external actor using "offensive agentic techniques" remains unexplained — the reporting doesn't establish that an AI agent did anything here, despite what that phrase invites you to imagine.

This is the blockchain-analysis firm whose customers include law enforcement, exchanges and compliance teams, so a full-user token reset plus multi-day product outages is a real operational event, not a footnote. It's also a tidy example of incident scope creep: "no evidence" on day one quietly becoming "everyone, actually" six days later.

Chainalysis hasn't said whether every data category was exposed for every user, how many users it has, what "live tokens" covers, or how passwords were hashed.

Apache OFBiz 24.09.08 fixes a pile of pre-auth RCEs

Apache's oss-security announcement on Tuesday catalogued a batch of OFBiz vulnerabilities, all fixed in the newly released 24.09.08, with everything before that version affected. The headliners are not subtle. CVE-2026-68524 is pre-authentication remote code execution through attacker-controlled screen definition loading. CVE-2026-73578 is an unauthenticated saved-view authorisation bypass with RCE impact. CVE-2026-83542 is unauthenticated RCE via login-page widget injection — yes, the login page. Then there's CVE-2026-105330: a hard-coded cryptographic key in JWT authentication that hands over unauthenticated administrative access. Rounding it out, the REST API's authentication is bypassed by default when the auth attribute is omitted, and the password-reset flow enables full account takeover. A solid afternoon's reading, if you enjoy despair.

OFBiz is an ERP framework, so it tends to sit near order, inventory and customer data. Several of these paths need no credentials at all, and the hard-coded JWT key is the kind of thing that makes you wonder how long it's been sitting there.

APT retry bug lets an on-path attacker swap a verified package for root

DSA 6607-1, republished to oss-security on Tuesday with Ubuntu's USN-8996-1 alongside, covers CVE-2026-109518 in APT — the thing that installs everything else on your Debian and Ubuntu boxes. The bug: APT can merge two package downloads after an HTTP redirect gives them the same URI, and improper file handling while resuming a download lets a man-in-the-middle attacker replace a package after verification. Result: arbitrary code execution as root. Ubuntu's notice spells out that this includes someone running a malicious mirror or proxy. The mitigating detail is that the attack needs redirects plus a non-zero Acquire::Retries setting — but retries default to 3 from Debian 12 and Ubuntu 22.04 LTS onwards, so current default installs are in scope while older releases aren't. Debian trixie is fixed in APT 3.0.4, Ubuntu has fixed builds for 26.10, 26.04 LTS, 24.04 LTS and 22.04 LTS, with fixes back to 14.04 via Ubuntu Pro or Legacy Support. Setting Acquire::Retries to 0 works as a stopgap.

Package managers are the trust anchor of the whole system, and this one undermines the verification step itself — the package passes the check, then gets swapped. The silver lining is the attacker has to already control the download path, and no real-world exploitation has been reported.

Debian's own list pages were erroring, so details here rest on the oss-security republication, Ubuntu's notice and the public advisory record.

DreamHost's File Manager bug spent two months harvesting FTP logins and webmail sessions

DreamHost disclosed an incident last week; the new California breach notice fills in what that actually meant. According to the notification quoted by BleepingComputer, an unauthorised party exploited a vulnerability in the web-based File Manager at files.dreamhost.com and spent roughly August to 23 October collecting FTP and SFTP usernames and passwords from people who logged in. The intruder may also have reached webmail systems, exposing session information — credentials, in some cases — for what DreamHost calls a limited number of users. TechCrunch reports the California filing says session tokens were taken, and that a spokesperson put affected customers at "well under 10%" of the company, which touts about 400,000 paying customers, without giving a firmer number. DreamHost says it blocked the access, fixed the vulnerability, killed affected sessions and invalidated passwords, and found no evidence control-panel logins or billing details were touched. The exploited vulnerability itself wasn't identified, though the company says it was unrelated to the recently disclosed PHP flaws.

Two months is a long dwell time for something harvesting hosting credentials, and stolen FTP logins plus live webmail sessions are the raw ingredients for site defacement, malware hosting and email compromise downstream. "Well under 10%" of 400,000 still leaves a lot of room.

The California filing wasn't read directly; the access window, credential scope and remediation claims are as reported by BleepingComputer and TechCrunch.

ChatGPT's Lockdown Mode had a popup-shaped escape hatch

Check Point Research has disclosed a bypass of ChatGPT's Lockdown Mode — the setting that's supposed to blunt prompt-injection exfiltration by cutting off live web access, remote images, connectors, Agent Mode and file downloads. The researchers found the HTML-preview sandbox shipped with allow-popups and allow-popups-to-escape-sandbox permissions, and that a nested iframe's contentWindow.open handed them an unrestricted window-opening function. In the Electron-based ChatGPT Classic desktop app, they say window.open worked with no user gesture at all, letting a malicious GPT silently send chat messages to an attacker-controlled server. They also demonstrated attacker-controlled windows good for fake-update and credential-phishing tricks. The limitations are worth keeping: browser-based malware delivery generally needed an initial user interaction, downloads in ChatGPT Classic still required a click, and this is a demonstrated vulnerability, not confirmed attacks in the wild. OpenAI told Check Point on 18 September that the sandbox permissions had been removed, and retesting confirmed the technique no longer works.

The whole point of Lockdown Mode is that cautious users and enterprises can turn it on and worry less about prompt injection leaking conversations. A bypass that defeats it silently in the desktop client is exactly the failure mode the feature exists to prevent — and it's a useful reminder that sandbox permission flags are load-bearing.

The fix predates this disclosure by several weeks, so the news is the research rather than a fresh patch; everything technical here is Check Point's account.