Binance confirms a breach with up to $5.8 billion in outflows in what may be the largest exchange theft on record, CISA adds GitLab and Cisco flaws to KEV after exploitation, Palo Alto Networks reports PHP2Shell attacks on exposed management interfaces, and hacktivists publish the work details of 41,000 DHS staff.
🔴 CRITICAL (4)
Binance confirms breach after up to $5.8 billion drained from hot wallets Under investigation
Binance says it is investigating a security breach after on-chain trackers put outflows at 5.5 to 5.8 billion dollars across three blockchains, which would make it the largest exchange theft on record. Tether and Circle froze roughly 1.8 to 2.4 billion dollars of stablecoins tied to the wallets, while the attacker swapped remaining freezable tokens for ether on decentralised exchanges. Bitcoin and ether already moved to unattributed wallets cannot be recalled, so recovery depends on the frozen stablecoins and exchange blacklists.
Sources CoinDesk|Decrypt|BleepingComputer|Decrypt
CISA adds GitLab CVE-2026-97229, CVSS 10.0, to KEV after attack on Alpine Linux server Confirmed
CISA added the critical GitLab flaw CVE-2026-97229 to its Known Exploited Vulnerabilities catalog on 29 October, giving federal agencies until 1 November to patch. The bug lets an authenticated user run code on self-managed GitLab Community and Enterprise servers through a crafted regular expression in a CI configuration, with the underlying fault in Ruby’s Onigmo regex engine. Alpine Linux disclosed that an attacker tried it against the project’s own GitLab server a week after fixes shipped, and said it found no evidence the attempt worked; any self-managed GitLab instance still unpatched should be treated as compromised.
Sources CISA|The Hacker News|Alpine Linux
Cisco patches exploited ASA and FTD flaw CVE-2026-76418 that can crash firewalls Confirmed
Cisco released fixes for CVE-2026-76418, an 8.6-rated flaw in Secure Firewall ASA and Secure Firewall Threat Defense software that lets an unauthenticated remote attacker force a reload of an internet-facing device. Cisco says the flaw has been exploited in the wild, and CISA added it to KEV on 29 October. It is the second exploited Cisco firewall VPN zero-day since August, so edge firewall and VPN appliances should be patched and checked for unexpected reloads.
Palo Alto Networks confirms PHP2Shell exploitation against exposed management interfaces Confirmed
Palo Alto Networks updated its advisory to state that active exploitation of the PHP2Shell flaws CVE-2026-109618 and CVE-2026-109619 has been observed against its management interface where that interface is reachable from the internet. CISA is separately warning agencies to fix a high-severity Linux kernel privilege escalation bug it says attackers chain with PHP2Shell to reach root after gaining code execution. Any PHP-based management plane exposed to the internet should be patched to the fixed builds now, and assumed compromised if it was reachable before patching.
Sources The Hacker News|Unit 42|BleepingComputer
🟠 HIGH (5)
Polymarket halts trading after third-party breach seeded crypto-draining script Under investigation
Polymarket paused trading and relayer actions on 29 October after attackers compromised a third-party service and used it to serve cryptocurrency-draining JavaScript to users. Roughly 2.5 million dollars in assets has been traced to a suspected attacker wallet, with one analysis putting about 2.65 million dollars in POL in that wallet. Users who transacted during the window should revoke token approvals and treat missing balances as stolen.
Sources BleepingComputer|Decrypt|CoinDesk
319 Chrome extensions with 13.4 million installs hijacked searches through tracking servers Confirmed
Researchers identified 319 Chrome extensions, with a combined 13.4 million reported installations, that redirected user searches through tracking servers to monetise Bing and Yahoo advertising. The extensions still returned plausible results, so most users had no reason to notice the interception. Remove any browser extension you do not actively use and audit the remainder against the published list.
Sources BleepingComputer
Linux kernel TCP flaw CVE-2026-98432 rated 9.8 has a working AI-written exploit Confirmed
A flaw in the Linux kernel’s TCP code, tracked as CVE-2026-98432, affects Ubuntu, Red Hat Enterprise Linux and other distributions and is rated 9.8 by the kernel’s own CVE team. Anthropic said on 27 October that one of its models wrote an exploit that lets an unauthenticated remote attacker run commands as root on exposed Linux servers. This is a different bug from the kernel flaw added to KEV on 28 October; apply distribution kernel updates and restrict exposed network services until patched.
Sources The Hacker News
Hacktivists publish work details of more than 41,000 DHS employees Under investigation
A hacktivist group has published names, job titles and phone numbers for more than 41,000 US Department of Homeland Security staff. SecurityWeek reports the data was allegedly taken from an Okta single sign-on system earlier this year. The records are work contact information rather than credentials, but they are directly usable for targeting and impersonation, so DHS-linked staff should expect convincing follow-on phishing.
Sources SecurityWeek|Decrypt
Telegram for Android call chain gave attackers code execution on answered calls Confirmed
NebuSec disclosed a Telegram for Android flaw that, chained with two other bugs, let any ordinary Telegram account take full control of a victim’s phone once the victim answered a voice call. Telegram fixed the issue in version 12.10.4, released in late September, and the researchers published details on 29 October. Update older Android builds now; exploit details are public and the attack needs no special access to the target.
Sources The Hacker News
🟡 MEDIUM (9)
Hugging Face crypto-mining campaign returns days after takedown, hijacking GPU servers Confirmed
A campaign that abuses Hugging Face models to plant cryptocurrency miners and backdoors returned days after a takedown, with one malicious model downloaded more than 28,000 times. The payload targets GPU servers, where mining is most profitable, and leaves persistent access behind. Separately, a fake Solana starter kit on GitHub is built to trick AI coding agents into installing a NetSupport RAT.
Sources BleepingComputer
Japan Airlines says Amadeus system breach exposed data of about 92,000 customers Confirmed
Japan Airlines disclosed that a breach at Amadeus, which runs its Passenger Service System, exposed the data of roughly 92,000 customers. The incident sits with the booking platform rather than JAL’s own systems, the pattern behind several airline breaches this year. Affected customers should expect targeted booking and refund scams that quote real itinerary details.
Sources BleepingComputer
Anthropic-themed phishing emails slip past leading secure email gateways Confirmed
A phishing campaign impersonating Anthropic is reaching inboxes despite obvious red flags and is bypassing several industry-leading secure email gateway products, according to DarkReading. The emails lean on interest in AI assistants to drive credential entry and malware delivery. Treat vendor-branded AI notifications as untrusted by default and verify through the vendor’s own site rather than the link in the message.
Sources DarkReading
Adobe, Veeam and Juniper patch critical vulnerabilities Confirmed
Adobe, Veeam and Juniper Networks each shipped fixes for critical flaws that could allow arbitrary code execution, authorisation bypass and privilege escalation. The vendors have not reported exploitation, but all three products sit in paths attackers value, including backup infrastructure. Apply the updates in the current patch cycle rather than waiting for the next maintenance window.
Sources SecurityWeek
ICS advisories: Hitachi Energy CoreTec and TRO600 flaws, Siemens Mendix SAML bypass Confirmed
CISA published advisories for Hitachi Energy CoreTec, which carries an OS command injection and a Dirty Frag write-what-where flaw rated 8.8, and for Hitachi Energy TRO600 series devices, where missing authentication and channel restrictions allow remote code execution. A third advisory covers a Siemens Mendix SAML authentication bypass, CVE-2026-89209, that lets an unauthenticated attacker replay a valid SAML assertion in IdP-initiated flows. Energy and manufacturing operators should apply the vendor updates and remove internet exposure from management interfaces.
26,400 valid cloud and AI credentials found in public Gitee repositories Confirmed
An independent researcher scanned 7.25 million public repositories on the Chinese code host Gitee and found 26,400 credentials that still worked when tested, mostly keys for cloud and AI services. Anyone who copied the code inherited usable access. This is the same class of exposure that keeps appearing on GitHub; secret scanning and key rotation remain the only reliable controls.
Sources The Hacker News
Hackers steal 2.6 million records from Sony's Japanese ISP So-net Confirmed
Attackers took 2.6 million records covering So-net and Plala customers in Japan, including 54,401 email account passwords and 16 bank account records. The stolen passwords make credential stuffing against linked accounts the immediate risk. Affected customers should rotate the exposed passwords and enable phishing-resistant multi-factor authentication where offered.
Sources SecurityWeek
Shadow AI relay resells Codex and Claude Code with sandboxes disabled Confirmed
An opaque relay service is reselling OpenAI Codex and Anthropic Claude Code access into mainland China and other regions where the products are not officially sold, with key security features turned off. Users of the relay get coding agents with reduced isolation, and any code or credentials passed through it should be considered exposed to a third party. Organisations should block unsanctioned agent relays at the network edge and enforce vendor-direct authentication.
Sources DarkReading
Talos documents NIK C2, an AI-enabled command-and-control framework seen in the wild Confirmed
Cisco Talos reported a new command-and-control framework called NIK C2 that integrates AI features and has been observed in real intrusions rather than only in research. The framework handles operator tasking and traffic management, which lowers the skill needed to run an intrusion. Detection should focus on beacon patterns and TLS fingerprints rather than known tool names, since the framework is new to most signature sets.
Sources Cisco Talos
⚪ WATCH (9)
Fake crypto app on Apple App Store linked to theft from 27 Solana wallets Confirmed
A fake cryptocurrency trading app called Padre Terminal, previously available on the Apple App Store, has been linked to the theft of about 23,000 dollars from 27 Solana wallets. The losses are small, but the case shows that store review does not guarantee safety for wallet-handling apps. Prefer wallet software distributed by the wallet vendor itself and verify the publisher before granting signing permissions.
Sources BleepingComputer
Another ClickFix campaign pushes fake fix instructions to visitors Confirmed
SANS Internet Storm Center documented a further ClickFix campaign that convinces visitors a site error requires them to paste a command into the Windows run dialog. The technique continues to work because it turns the user into the execution step, bypassing many email and web controls. User awareness and blocking clipboard-to-run behaviour remain the practical defences.
Sources SANS ISC
Microsoft tracks uneven AI adoption across North Korean threat actor tradecraft Confirmed
Microsoft Threat Intelligence described how North Korean actors are adopting AI at different rates across their operations, from phishing text generation to tooling support. The report treats adoption as incremental rather than transformative, with tradecraft and infrastructure still doing most of the work. Defenders should expect faster iteration on lures and code, not new capabilities overnight.
Sources Microsoft Security Blog
Google says Gemini 4 Carbon leads vulnerability discovery benchmarks, opens it to vetted defenders Confirmed
Google introduced Gemini 4 Carbon, a frontier model it says scores 88.2 percent on SEC-bench Pro and leads current models at finding, exploiting and patching vulnerabilities. Access is being rolled out first to vetted defenders through the Fairwind Program rather than to the general public. The same capability advantage will not hold for long, which is the argument for shortening patch timelines now.
Sources SecurityWeek|The Hacker News
New quantum model cuts hardware needed to attack bitcoin keys by 17 times Researcher disclosure
Alice and Bob published a model that reduces the hardware required to run a quantum attack against bitcoin’s elliptic curve keys by roughly 17 times, bringing the theoretical attack closer to practical scale. No such attack has been demonstrated, and the work is a planning input rather than a break. It strengthens the case for tracking post-quantum migration timelines for long-lived key material.
Sources CoinDesk
CrowdStrike details prompt injection techniques aimed at agent planning stages Researcher disclosure
CrowdStrike described new prompt injection techniques that target an AI agent’s planning stage rather than its final output, so the malicious instruction shapes the whole task decomposition. Attacks of this shape are harder to catch with output filters because each individual step looks reasonable. Logging full agent plans and constraining tool permissions remain the strongest mitigations.
Sources CrowdStrike
Talos IR trends: help desk vishing rises as ransomware share of cases grows Confirmed
Cisco Talos reported that phishing remained the leading initial access technique in the third quarter, with actors increasingly impersonating IT help desks in voice phishing calls. The share of ransomware and pre-ransomware cases rose, and finance and public administration tied as the most targeted industries. Help desk identity verification is the control that most directly addresses the observed pattern.
Sources Cisco Talos
Check Point reads Cohere's Qwen incident as model jaggedness, not sabotage Researcher analysis
Check Point analysed Cohere’s report that Qwen3.8-Flash-Next produced a shell command that inspected the host while writing a detection rule, and concluded the behaviour looks more like uneven model capability than deliberate action. The case is a useful example of agent output that is plausible, on task and still unsafe to execute. Run agent-generated commands in a sandbox and review them before execution.
Sources Check Point Research
GitHub previews an install-time firewall for npm packages Vendor announcement
GitHub launched a preview package firewall that inspects metadata requests and package downloads for signs of malicious activity and blocks installation before code runs. The control targets the install step, where most npm supply-chain attacks execute. Teams should trial it alongside lockfile review and dependency pinning rather than as a replacement for either.
Sources The Hacker News
Methodology: compiled from vendor advisories, government feeds (CISA KEV, MSRC), security news sources and on-chain/security-firm alerts. Grouping: one incident, one entry, with every source cited. Unconfirmed reports are labeled. Crypto items are incident reporting, not investment advice. Corrections from prior digests are noted at the top when applicable.