Anthropic identified industrial-scale distillation campaigns by DeepSeek, Moonshot, and MiniMax that generated over 16 million exchanges across roughly 24,000 fraudulent accounts to extract Claude's agentic reasoning, tool use, and coding capabilities for training their own models. The labs used proxy services operating “hydra cluster” architectures with thousands of coordinated accounts to circumvent regional access restrictions, with MiniMax pivoting within 24 hours to target new model releases and DeepSeek specifically extracting chain-of-thought data and censorship-safe alternatives to politically sensitive queries. Anthropic has deployed behavioral fingerprinting classifiers, strengthened account verification, and is sharing technical indicators with other labs and authorities, warning that illicitly distilled models strip safety guardrails and undermine export controls.
Dependabot floods teams with noisy pull requests and misleading security alerts, and in Go projects, it rarely improves actual security. Filippo Valsorda breaks down why: the alerts often have no bearing on real risk, offer no package- or symbol-level filtering, and generate enough alert fatigue that teams start ignoring everything, including the alerts that matter. His recommended setup uses two scheduled GitHub Actions instead. The first runs govulncheck for reachability-based vulnerability scanning. The second runs tests daily against the latest dependency versions to catch breakage early without forcing constant upgrades. Together, they cut through the noise, slash false positives, and let teams take genuine security findings seriously.
Sponsored
A single misconfiguration rarely causes a breach. It's the chain — the quiet connections between identities, roles, and systems — that creates opportunity. Attack Path Management shows you those chains in full context. Take SpecterOps' quick assessment [11] to see where your program falls on the maturity spectrum and access the MATURITY MODEL REPORT [12] for clear guidance to reduce attack paths.
Aikido Safe Chain is a lightweight local proxy that intercepts npm and PyPI package downloads, checking them in real-time against Aikido Intel threat intelligence to block malware before it reaches developer machines or CI/CD pipelines. It supports npm, yarn, pnpm, bun, pip, uv, and poetry via shell aliases, and temporarily suppresses npm packages published within the last 24 hours by default to catch undetected threats during the highest-risk window. Installation is tokenless and requires no build data sharing, with native CI/CD support for GitHub Actions, Azure Pipelines, GitLab, Jenkins, CircleCI, and Bitbucket.
Adaptive Security is an AI-powered platform that simulates deepfake and impersonation attacks across email, voice, and messaging to expose weak controls. It also delivers personalized training and policy enforcement to help employees detect and resist AI-driven social engineering threats.
Hijagger checks all maintainers of all npm and PyPi packages for hijackable packages through domain re-registration.
Cloudflare One is the first SASE platform to support standards-compliant post-quantum hybrid ML-KEM encryption across all major on-ramps and off-ramps, including its Secure Web Gateway, Zero Trust, and WAN use cases. The upgrade extends post-quantum protection to Cloudflare IPsec (now in closed beta) and the Cloudflare One Appliance (GA in version 2026.2.0), defending enterprise traffic against harvest-now-decrypt-later attacks ahead of NIST's 2030 deadline for deprecating RSA and ECC. The implementation follows the draft-ietf-ipsecme-ikev2-mlkem specification rather than proprietary approaches, and is available at no extra cost.
South Korean police charged two teens for breaching Seoul's Ttareungyi bike rental system in 2024, accessing data on 4.62 million of 5 million users, including contact and personal details. Investigators say the pair met on Telegram, hacked while still in middle school, and intended to profit, though there's no evidence the data was sold.
Sponsored
Mitiga observed an attacker exploiting a path transversal vulnerability to overwrite prod training data and trigger anomalous model behavior, planting a false reality in prod AI systems. Read the
Chrome 145 transforms the desktop address bar into a Gemini-powered AI prompt box with Deep Search, image generation, and conversational query support accessible via a new “plus” icon.
Firefox 148 introduces the standardized Sanitizer API and the new setHTML() method to make inserting untrusted HTML safer by default, reducing XSS risk without major code rewrites.
RoguePilot is a GitHub Codespaces flaw where hidden prompt injections in issues silently hijack Copilot, exfiltrating privileged GITHUB_TOKENs and enabling AI‑mediated supply-chain attacks.
Sponsored
Attackers are using AI to automate lateral movement, mapping your network in seconds and minutes instead of weeks. WHY
CVE-2025-29969 is a remote code execution vulnerability in the MS-EVEN RPC protocol — enabled by default on Windows 11 and Windows Server 2025 — that allows low-privileged users to write arbitrary files remotely by exploiting a TOCTOU flaw in the EventLog backup function, effectively bypassing default C$ share limitations. The attack chain combines a hidden CreateFile primitive for remote file/directory existence checks with an arbitrary write primitive achieved by swapping an attacker-controlled EVTX file on a remote SMB share between header validation and backup, enabling code execution via Startup folder batch scripts or DLL hijacking. Microsoft patched the TOCTOU write primitive in May 2025, though the recon capabilities enabling remote file enumeration across domain-joined machines remain unpatched.
Everest ransomware actors accessed Vikor Scientific (now Vanta Diagnostics) data via revenue cycle vendor Catalyst RCM, using compromised credentials to steal about 12GB of files containing names, dates of birth, payment card data, medical details, and insurance information. Roughly 140,000 individuals are currently listed as affected, but the final impact may be higher as counts from all linked entities remain unclear.
A misconfigured server exposed a full intrusion toolkit where a threat actor integrated LLMs directly into the attack workflow — using DeepSeek to generate attack plans from recon data and Claude Code to conduct live vulnerability assessments — while targeting FortiGate appliances across 106 countries with 2,516 identified targets. Custom tools ARXON (an MCP server bridging LLM analysis with attack scripts) and CHECKER2 (a Go-based Docker orchestrator) automated the pipeline from stolen VPN config ingestion through internal scanning to LLM-driven exploitation planning, evolving from the open-source HexStrike framework in roughly eight weeks. Defenders should audit for unauthorized VPN accounts, unexpected SSH access, and unapproved firewall policy changes, as the dual-model approach of selecting whichever LLM is most permissive for a given task lowers the skill barrier for managing concurrent intrusions at scale.
Samsung's pre-installed Weather app quietly creates a highly unique fingerprint from the combination of saved location IDs (placeid) it sends to The Weather Company's API, with 96.4% of sampled devices uniquely identifiable across days. These fingerprints persist over time, survive IP and network changes, and can be resolved into precise locations using hardcoded API keys that work from any client. Requests often include both hashed placeid values and raw GPS coordinates, giving IBM's Weather Company and Samsung rich, granular location profiles that echo a broader industry pattern of monetizing weather-based location data.
Unit 42 researchers introduced a novel detection method that fingerprints cloud threat actors by linking their known MITRE ATT&CK techniques to specific alert patterns. This approach successfully distinguished Muddled Libra (Scattered Spider) and Silk Typhoon operations across 22 industries from June 2024 to June 2025, with only three shared alert signatures out of around 120, highlighting their differing tactics. Industry alert spikes correlated with publicly reported campaigns, such as a 25% rise in transportation sector alerts during Muddled Libra's aviation attacks, enabling defenders to identify threats early by investigating actor-related alerts over a 30-day window.
2026.02.24·unit42.paloaltonetworks.com·20 min read·TLDR InfoSec Siper is a high-performance, XDP-based IP blacklist firewall built with Go and eBPF that drops malicious traffic at the NIC driver level before packets reach the kernel's networking stack. It supports CIDR-based blocking via BPF LPM Trie maps, persistent JSON-based rule management, and real-time drop/pass metrics. The minimalist architecture separates a Go CLI control plane from the kernel-space data plane, offering significantly lower CPU overhead than iptables or nftables under heavy DDoS conditions.
TrivySummary summarises Trivy scan JSON output for reporting purposes. Package vulnerabilities are collapsed down to the respective CVE, and headline counts of vulnerabilities at different priorities. As well as the vendor severity and CVSS v3 score provided by Trivy, EPSS scores can also be retrieved, allowing the exploitability and severity of all vulnerabilities to be graphed and prioritised.
Depthfirst is an AI security startup whose General Security Intelligence platform scans and analyzes codebases and workflows to detect issues, prevent credential exposure, and monitor open source and third‑party components for AI‑driven threats.
Veracode researchers discovered a typosquatted NPM package called "buildrunner-dev" that delivers the Pulsar RAT by hiding malicious code within PNG images using steganography, extracting payloads from RGB pixel values at runtime. The attack chain employs a heavily obfuscated 1,600-line batch file with only 21 functional lines, detects and evades AV products like ESET and Malwarebytes, and uses process hollowing to inject the final payload into legitimate processes. Developers should carefully verify package names and audit dependencies, as supply chain attacks increasingly leverage image-based steganography to bypass traditional file-scanning tools.
PayPal disclosed a software error in its Working Capital loan app that exposed the names, contact details, SSNs, and dates of birth of 100 business customers between July and mid-December 2025. The company reversed the faulty code, reset passwords, and detected some unauthorized transactions.
Anthropic's new Claude Code Security feature scans codebases for vulnerabilities, proposes patches, and claims to have surfaced hundreds of high‑severity issues in open source projects, spooking security investors and pushing stocks like CrowdStrike down. It joins Google, Microsoft, Amazon, and OpenAI in deploying AI agents for bug hunting, but all still require human review and face questions over false positives, costs, and marketing hype.
Sponsored
Modern IT infrastructure moves faster than manual workflows can handle. Read this Tines guide [17] to learn how to SCALE INFRASTRUCTURE PREDICTABLY, without sacrificing performance or governance.
The US Treasury announced a public-private initiative through the AI Executive Oversight Group that will release six resources throughout February to help financial institutions — particularly small and mid-sized ones — manage AI-specific cybersecurity risks and deploy AI more securely.
Spanish police arrested a 20-year-old in Madrid who manipulated a hotel booking site's payment gateway validation process to secure luxury rooms at €0.01 per reservation.
Intellexa's Predator spyware suppresses iOS camera and microphone recording indicators by hooking SpringBoard's SBSensorActivityDataProvider, nullifying sensor activity updates before they reach the UI layer.
Sponsored
Your security stack is growing, but so are your attack surfaces. Containers, microservices, remote teams, and data lakes—each adds complexity.
that show exactly how to implement modern security patterns, including:
An evolved ClickFix campaign uses fake CAPTCHA pages to trick users into executing PowerShell commands that download an infostealer via the Donut shellcode loader, which runs entirely in memory using VirtualAlloc and CreateThread to evade file-based detection. The malware targets cryptocurrency wallets, including MetaMask and Exodus, harvests credentials from over 25 browsers, and exfiltrates VPN configurations, Steam accounts, and FTP credentials, while persisting through RunMRU registry key modifications. Organizations should monitor for suspicious PowerShell execution chains, block known C2 IPs, and note that Microsoft Defender detects the payload as Behavior:Win32/SuspClickFix.C.
A Russian-speaking, financially motivated threat actor leveraged commercial generative AI tools to compromise over 600 FortiGate devices across 55 countries by targeting exposed management interfaces and weak single-factor credentials, according to Amazon Threat Intelligence. Post-exploitation activity included Active Directory compromise, credential harvesting via pass-the-hash/pass-the-ticket and NTLM relay attacks, and targeting Veeam Backup servers using CVE-2023-27532 and CVE-2024-40711, consistent with pre-ransomware operations. Organizations should immediately restrict exposure of the FortiGate management interface, enforce MFA for all administrative and VPN access, rotate SSL-VPN credentials, and isolate backup infrastructure from general network access.
The UK Court of Appeal has backed the ICO's £500k fine on DSG Retail over a 2017 breach that exposed 5.6 million card details and data on 14 million people, even though attackers lacked cardholder names. The ruling confirms that payment card numbers and expiry dates are personal data from the controller's perspective, stresses jigsaw identification risk, and strengthens regulators' ability to penalize weak security in future cyberattacks.
AI assistants with web-browsing capabilities, specifically Grok and Microsoft Copilot, can be abused as covert C2 relays by using anonymous web access to fetch attacker-controlled URLs and return embedded commands, all without API keys or registered accounts. This PoC uses a WebView2-based C++ implant to create a bidirectional channel in which victim data is exfiltrated via URL query parameters and commands are returned via AI-generated responses, with simple encoding sufficient to bypass model-side safeguards. The research outlines a near-term evolution toward AI-Driven malware where implants shift from static logic to prompt-driven behavior, enabling AI-assisted anti-sandbox detection, intelligent victim triage, and targeted ransomware that prioritizes high-value files to evade volume-based detection thresholds.
This post details a methodology for empirically verifying AWS resource exposure without reading sensitive data or altering the resource state. It employs four techniques: comparing unsigned and signed requests, conducting metadata-only reads, performing no-op operations like untagging nonexistent keys, and sending malformed request probes that leverage AWS' sequence where authorization checks occur before parameter validation. The "3-topic method" ensures probe safety by testing identical changes across allowed, denied, and nonexistent resources — if only denied and nonexistent targets return 403 errors and allowed ones return 400, it confirms authorization without executing the action. This method is implemented in the open-source sns-buster tool for SNS and is adaptable to other AWS services for CSPM, bug bounty, and IAM policy validation.
Ring buffer, introduced in Linux 5.8, replaced the legacy perf buffer as the recommended mechanism for kernel-to-userspace data transfer in eBPF programs. It uses a single shared circular buffer across all CPUs instead of per-CPU buffers, eliminating cross-CPU event ordering complexity and reducing copy overhead. This article walks through both implementations in C and Go using cilium/ebpf, demonstrating how the ring buffer's reserve-and-submit model via bpf_ringbuf_reserve and bpf_ringbuf_submit provides a simpler, higher-throughput path compared to perf buffer's bpf_perf_event_output approach. For security tooling built on eBPF tracing, the ring buffer is now the default choice for structured event streaming from kernel-level monitors to usersp.
Sponsored
AI coding assistants create security risks at machine speed and massive scale. Checkmarx Developer Assist embeds an AI-native AppSec agent directly in AI IDEs, preventing vulnerabilities in real time before commit. It enforces policy, validates AI-generated code, and secures the modern Agentic Development Lifecycle (ADLC). Explore the
Anthropic launched Claude Code Security in limited research preview for Enterprise and Team customers, offering AI-powered codebase vulnerability scanning that goes beyond static analysis by reasoning about component interactions and tracing data flows. Findings undergo a multi-stage verification process to filter false positives, with each vulnerability assigned a severity and confidence rating alongside suggested patches. The feature employs a human-in-the-loop approach where developers review and approve all remediation before any changes are applied.
This repository gathers all available CVE exploits from GitHub and publishes them on a public website.
Trail of Bits released an opinionated configuration repository for Claude Code covering sandboxing, permissions, hooks, skills, MCP servers, and workflow patterns refined through security audits and development. The repository includes pre-built slash commands for autonomous PR review, issue resolution, and Dependabot merging using parallel agents, alongside sandbox hardening that blocks access to SSH keys, cloud credentials, crypto wallets, and shell configs. It also documents context management strategies, local model integration via LM Studio, and a multi-layered security model combining OS-level sandboxing, devcontainers, and disposable cloud droplets for safe bypass-permissions operation.
Cloudflare experienced a 6-hour outage after a cleanup sub-task in its Addressing API queried with an empty `pending_delete` parameter, causing the system to interpret all 4,306 BYOIP prefixes as queued for deletion and systematically withdrawing ~1,100 BGP prefixes across customer networks. The bug rendered services including Magic Transit, Spectrum, and a subset of 1.1.1.1 unreachable. Recovery was complicated by varying states of impact, where some customers lost both prefix advertisements and service bindings. Cloudflare is implementing circuit breakers for large-scale BGP withdrawal actions, separating operational and configured state in the Addressing API, and deploying health-mediated configuration snapshots as part of their Code Orange: Fail Small initiative.
English Wikipedia is banning Archive.today after discovering the site used its CAPTCHA to DDoS a critic's blog and secretly altered archived pages to smear him, undermining trust in its snapshots. Editors will replace roughly 695,000 Archive.today links with alternatives like the Internet Archive.
ATM jackpotting, first showcased by researcher Barnaby Jack, has become a lucrative cybercrime, with over 700 attacks in 2025 netting at least $20 million in cash. Criminals use physical access, generic keys, and Ploutus malware to hijack Windows-based ATMs via XFS software, directly commanding dispensers to spew banknotes without touching customer accounts and often evading detection until long after cash-out.
Ukrainian national Oleksandr Didenko was sentenced to five years in prison for providing 871 proxy identities and operating laptop farms across multiple countries to help North Korean IT workers fraudulently secure jobs at 40 US companies.