标签：#hr

IPMI/BMC Authentication Hash Leak: Stealing Out-of-Band Server Management via Offline Cracking 1. Basic Information Article Title : Over 24,000 exposed server BMCs leak password hash via decades-old flaw Publisher : BleepingComputer Publication Date : July 28, 2026 Original Source : https://www.bleepingcomputer.com/news/security/over-24-000-exposed-server-bmcs-leak-password-hash-via-decades-old-flaw/ Related Sources : Lava research (linked in original), CVE-2013-4786 Related Entities : CVE-2013-4786, IPMI 2.0, BMC, Supermicro, HPE iLO 4, UDP/623, GPU/AI servers Severity : High 2. Executive Summary Attackers can send an authentication request to exposed IPMI/BMC interfaces on the internet and receive a password-derived response. They can crack this response offline using GPUs to gain server management privileges that sit below the operating system. A recent scan found 24,650 devices leaking authentication material, with 2,340 using weak, dictionary-based administrator passwords. 3. Attack Flow Scan the internet for IPMI 2.0 services on UDP/623. Start an RMCP+/RAKP authentication handshake with the BMC. Obtain the password-derived authentication response without credentials. Perform offline cracking using dictionaries, default ADMIN users, and vendor-generated password rules. Log in to the BMC using the recovered password. Control power, virtual media, firmware, and low-level settings. Reuse the same credentials on other management interfaces. Inference : Use virtual media or boot settings to compromise the OS, establish persistence, and move laterally across the management network. 4. Attacker Position and Execution Location Attackers connect to UDP/623 from the internet or a reachable management network. Hash cracking happens on the attacker's GPU/CPU, avoiding failed login alerts on the target system. Post-compromise actions run on the BMC processor, which is independent of the OS. 5. Visibility for Victims and Administrators The OS, EDR, and host logs may not show

Employees of OpenAI and Anthropic, as well as Google, Meta, Thinking Machines, Microsoft, Mistral, and other leading AI labs, have written a statement to the US government supporting a potential slowdown of sorts for frontier AI development - or at least a speed-up of global coordinated governance efforts. "Al could help create a dramatically better […]

It's earnings season, and investors got an unpleasant surprise from Google: an increase on its spending estimate, to as much as $205 billion - from the last quarter's projection of up to $190 billion. Even the lower end of Google's new projected range - $195 billion - is much more than the company had previously […]

Anthropic and OpenAI employees are expected to give generously after their companies go public. “It’s going to be a wild ride,” says one nonprofit leader.

I have been building VoxelDraft , a voxel editor that runs entirely in the browser without an account or installation. The editor supports block painting, layers, keyframe animation, GIF recording, local projects, and exports for OBJ/MTL, GLB, VOX, Minecraft Schematic, and Roblox RBXL. This post covers the architecture choices that kept those features manageable. Keep edit data serializable The editable model is an array of plain voxel records rather than a collection of Three.js objects: type VoxelData = { position : [ number , number , number ] color : string layerId ?: string } That decision makes JSON backups, local persistence, undo/redo snapshots, sharing, and format conversion much simpler. Three.js objects are derived render state, not the source of truth. Render repeated cubes with InstancedMesh Creating one mesh and one React component per cube becomes expensive as a model grows. VoxelDraft uses THREE.InstancedMesh where geometry and material can be shared. Each voxel contributes a transform matrix. Pointer intersections return the instanced mesh and instance ID, which can be mapped back to the editable voxel record. There are tradeoffs. Per-voxel colors need instance colors or grouping by material, and changing a single block still requires carefully updating the instance buffers. The reduction in draw calls is worth that complexity. Make exporters independent from UI The format exporters accept voxel records and produce a Blob . The UI is only responsible for validation and triggering a download. const blob = exportToVOX ( voxels ) const url = URL . createObjectURL ( blob ) VOX, Minecraft Schematic, and RBXL are generated directly. For GLB, the app builds a temporary Three.js scene and sends it to GLTFExporter from three-stdlib . Keeping binary generation separate from React event handlers makes exporters easier to test and reuse. Move GIF encoding off the main thread VoxelDraft records both animation output and modeling timelapses. GIF encoding can easi

Dysphoria: A 200k-Device Botnet Using Blockchain Name Resolution and Infected Device Relays 1. Basic Information Article Title : New Dysphoria DDoS botnet spreads to 200k devices worldwide Publisher : BleepingComputer Publication Date : 2026-07-27 Original Source : https://www.bleepingcomputer.com/news/security/new-dysphoria-ddos-botnet-spreads-to-200k-devices-worldwide/ Primary Source : https://blog.xlab.qianxin.com/dysphoria/ Related Entities : Dysphoria, jackskid, fbot, ENS, SNS, UPnP, Telnet, SSH, IoT/embedded Linux Related CVEs : CVE-2013-3307, CVE-2016-20016, CVE-2017-17215, CVE-2017-5259, CVE-2018-14558, CVE-2020-25499, CVE-2020-8515, CVE-2022-35733, CVE-2025-9528, CVE-2025-28137, CVE-2025-34152, CVE-2025-55182, and others Severity : High Main IOCs : burrberry.eth , ukranianhorseriding.eth , 24carnforth2merseyside.sol , login.trees4sale.net , c2.saintpetersburgresident.ru 2. Summary This is a large-scale Linux botnet. It breaks into devices using weak Telnet/SSH passwords and various IoT RCE vulnerabilities. It gets multi-stage C2 servers from ENS/SNS records. It turns infected devices into relay nodes or DDoS bots by opening 155 ports using UPnP. 3. Attack Flow Infection and DDoS Chain Attackers compromise routers, gateways, and cameras using weak Telnet/SSH credentials or known RCE vulnerabilities. The malware runs a Linux binary. It hides its process name as libdalvikengine.so . It decrypts strings using modified RC4 (with added LCG/LFSR). It reads TXT and custom records from ENS/SNS, converts fake IPv6 formats, and gets distributor node IP addresses. It gets a list of infected device relays from http://<node>:9000/nodes?key=meowmeowmeow . It connects to the C2 server using a fixed 78-byte login and heartbeat. It executes DDoS commands that include duration, attack type, target, and flag. Relay-Only Chain A relay variant without DDoS features searches for UPnP-enabled gateways. It creates 155 port forwards on the WAN side and listens on its own device. It

Public Wi‑Fi DNS Poisoning: Hijacking Microsoft 365 Sessions of Business Travelers 1. Basic Information Article Title : Hacked Public Wi-Fi Gateways Used to Harvest Corporate Credentials Source : SecurityWeek Publication Date : 2026-07-27 Original Article : https://www.securityweek.com/hacked-public-wi-fi-gateways-used-to-harvest-corporate-credentials/ Primary Source : https://reliaquest.com/blog/threat-spotlight-dns-poisoning-tactics-expand-to-hospitality/ Related Entities : Microsoft 365, Entra ID device-code flow, DNS poisoning, AiTM, WPAD, captive portal, FrostArmada, APT28 (attribution unconfirmed) Severity : High IOCs : 38.146.28.75 , 31.57.243.154 , 104.194.159.150 , m365-owa.com , owa-ms365.com , ms365-device.com , ms365-live.com 2. Executive Summary This attack compromises Wi-Fi gateways at locations like hotels to forge DNS responses. It does not send phishing emails. Instead, it tricks users into visiting fake Microsoft pages, using WPAD proxies, and entering device-code authentication. This allows attackers to steal MFA-authenticated Microsoft 365 sessions. 3. Attack Flow Chain A: DNS Redirection and Credential Theft Attackers gain administrative access to captive portal gateways in hotels or conference centers. Hypothesis (Primary source confidence: Low to Medium) : Public SSH/SNMP/Web management interfaces and weak, reused administrator credentials are used for initial access. The gateway returns fake responses for DHCP-assigned DNS or passing DNS, resolving normal domains to attacker IPs. Users are redirected to fake Microsoft pages to collect credentials and OAuth information. Attackers log into Microsoft 365 using the stolen credentials and tokens. Chain B: WPAD The gateway controls wpad name resolution or DHCP option 252. The device downloads the attacker's PAC file. Windows app and browser traffic pass through the attacker's proxy. Investigations observed attempts, but successful execution is unconfirmed. Chain C: Device-Code Flow The fake page sh

FortiOS CVE-2025-68686: Bypass of Symlink Persistence Mitigation for Already Compromised Devices 1. Basic Information Article Name : CISA Adds Two Known Exploited Vulnerabilities to Catalog Source : CISA Publication Date : July 27, 2026 Original Link : https://www.cisa.gov/news-events/alerts/2026/07/27/cisa-adds-two-known-exploited-vulnerabilities-catalog Related Sources : https://fortiguard.fortinet.com/psirt/FG-IR-25-934 https://nvd.nist.gov/vuln/detail/CVE-2025-68686 Related Entities : CVE-2025-68686, FortiOS, SSL-VPN, symlink persistence, CISA KEV Severity : Critical 2. Summary This is an actively exploited vulnerability. An attacker who has already compromised the FortiOS file system via another vulnerability can use a crafted HTTP request to bypass symlink persistence mitigations. This allows access to sensitive files through the SSL-VPN web interface even after an upgrade. 3. Attack Flow An attacker compromises FortiOS down to the file system level using another vulnerability or path. The attacker places a symlink pointing to out-of-bounds system files into the SSL-VPN related area. An administrator applies standard patches or firmware updates, but the malicious symlink or its recreation path remains. The attacker bypasses the mitigation using a crafted HTTP request via CVE-2025-68686. The attacker may read sensitive files such as configurations, credentials, and keys from the SSL-VPN web interface. The attacker may use the stolen information to continue VPN authentication, administrative access, and internal intrusion. 4. Attacker Position and Execution Location Initial compromise and symlink placement happen on the FortiGate/FortiOS device. Mitigation bypass and file reading occur externally via the SSL-VPN HTTP(S) interface. This CVE alone does not provide initial file system access to uncompromised devices. 5. Visibility for Victims and Administrators The device may look updated, but unauthorized symlinks can remain. Crafted HTTP requests to the SSL-VPN w

Certighost CVE-2026-54121: Low-Privilege Users Impersonate a DC via AD CS 1. Basic Information Article Title : New Certighost PoC exploit lets attackers hijack Windows domains Source : BleepingComputer Published Date : 2026-07-27 Original Article : https://www.bleepingcomputer.com/news/security/new-certighost-poc-exploit-lets-attackers-hijack-windows-domains/ Primary Source : https://gist.github.com/H0j3n/a5ef2609b5f2944ac2390a191a534c26 Related Entities : CVE-2026-54121, Certighost, Microsoft AD CS, Enterprise CA, PKINIT, Kerberos, DCSync, Impacket Severity : High 2. One-Sentence Summary A public PoC allows a low-privilege domain user to point AD CS "chase" references to an attacker host, trick the CA into trusting fake LDAP/LSA responses to get a Domain Controller certificate, and then take over the entire domain using PKINIT and DCSync. 3. Attack Flow An attacker connects to LDAP with a low-privilege domain account and lists CAs, DCs, and Domain SIDs/GUIDs. The attacker uses the default ms-DS-MachineAccountQuota=10 setting to create a computer account and register an SPN. The attacker starts fake LDAP and LSA services on their host. The attacker specifies their host in the certificate request's cdc (Client DC) field and the target DC in the rmd (Remote Domain) field. The CA connects to the attacker-specified cdc via SMB/LDAP to chase the reference. The attacker's service relays the CA's authentication challenge to the real DC and returns the target DC's objectSid and dNSHostName . The vulnerable CA treats the returned data as authoritative and issues a certificate containing the target DC's SID and DNS name. The attacker uses PKINIT to get a Kerberos TGT as the DC and saves it to a .ccache file. The attacker performs a DCSync using the DC account's replication rights to steal secrets, including krbtgt . 4. Attacker Position and Execution Location The attacker needs a host inside the domain with valid authentication that can reach the CA, DC, and SMB/LDAP ports. T

MedusaHVNC: Remote Control of Logged-in Browsers on Hidden Windows Desktops 1. Basic Information Article Title : MedusaHVNC Malware Uses Hidden Windows Desktops to Evade Detection Publisher : SecurityWeek Publication Date : 2026-07-27 Original Article : https://www.securityweek.com/medusahvnc-malware-uses-hidden-windows-desktops-to-evade-detection/ Primary Source : https://www.blackfog.com/medusahvnc-a-hidden-desktop/ Related Entities : MedusaHVNC, MaaS, HVNC, JScript, AutoIt, charmap.exe , ChaCha20, Chrome/Edge/Firefox Severity : High IOCs : 51.89.204.28:4444 , %TEMP%\Nx2981Okkr2\ , AFLlvOscPj.bat , zorsxklxfehdoals 2. Executive Summary This is a RAT that uses JScript, AutoIt, charmap.exe injection, and multi-layer decryption to open logged-in browsers on a hidden Windows desktop invisible to the user. It controls existing cookies and sessions using the victim device's original IP address. 3. Attack Flow wscript.exe runs an obfuscated JScript file and waits for 7,584 milliseconds. It extracts the AutoIt execution system, configuration, and encrypted payload to %TEMP%\Nx2981Okkr2\ . It places AFLlvOscPj.bat in the Startup folder for persistence. AutoIt decrypts the payload using single-byte XOR 0xAE . It starts the legitimate C:\Windows\System32\charmap.exe and injects the loader into it. It decrypts the final x64 PE file using 16-byte repeating XOR and ChaCha20. It makes a custom TCP connection to 51.89.204.28:4444 . It creates another hidden desktop and launches Chrome, Edge, or Firefox. It captures screens using BitBlt and PrintWindow , sends inputs using SendInput , and moves data using the Clipboard API. It uses cookies and sessions from existing browser profiles to control user accounts. 4. Attacker Position and Execution Location The attacker uses a MaaS operation panel and C2 server. The loader and HVNC run on the Windows device, and the browser runs on a separate desktop invisible to the user, but on the same device, IP, and profile. The initial delivery ve

VeloCloud Orchestrator CVE-2026-16812: Unauthenticated OS Command Injection Actively Exploited 1. Basic Information Article Title : Arista patches VeloCloud Orchestrator zero-day exploited in attacks Source : BleepingComputer (Primary Source: Arista Security Advisory 0144) Publication Date : 2026-07-27 Original URL : https://www.bleepingcomputer.com/news/security/arista-patches-velocloud-orchestrator-zero-day-exploited-in-attacks/ Related Sources : https://www.arista.com/en/support/advisories-notices/security-advisory/24364-security-advisory-0144 https://www.cisa.gov/news-events/alerts/2026/07/27/cisa-adds-two-known-exploited-vulnerabilities-catalog Related Entities : CVE-2026-16812, CWE-78, VeloCloud Orchestrator (VCO) On-Premises, VeloCloud Edge, CISA KEV Severity : Critical IOCs : 8.19.75.217 , 206.72.242.124 , 206.72.242.162 2. Summary This is a CVSS 10.0 vulnerability. It allows an unauthenticated attacker to access the Web UI of an internet-reachable on-premises VCO. The attacker can execute OS commands through internal-only functions. Active exploitation has been confirmed. 3. Attack Flow An attacker searches for a VCO Web interface. The attacker sends a crafted request without authentication to reach internal-only functions. The attacker executes commands on the VCO host via OS command injection. The attacker may access configurations, device lists, credentials, certificates, keys, and databases. The attacker proceeds to create files, export databases, create archives, perform outbound communications, and change management configurations. Inference : The attacker can abuse the authentication and configuration paths to Edge devices managed by the VCO, expanding the impact to the entire SD-WAN. 4. Attacker Position and Execution Location The attacker is an external host with network access to the VCO Web UI. The initial request targets the web layer. Commands execute on the on-premises VCO host. Hosted and Dedicated VCOs are already patched. VeloCloud Gateways

Anthropic founder and CEO Dario Amodei made his views clear about open-weight models and China's growing AI capabilities.

Meta on Monday said it is rolling out its Meta AI chatbot within Threads' DMs, giving users a way to chat with the AI assistant.

Steam Forum ClickFix: Fake Repair Commands Lead to XMRig SYSTEM Persistence 1. Basic Information Article Title : Steam forum ClickFix attacks infect gamers with XMRig cryptominers Publisher : BleepingComputer Publication Date : July 25, 2026, 18:37 EST / July 26, 2026, 07:37 JST Original Source : https://www.bleepingcomputer.com/news/security/steam-forum-clickfix-attacks-infect-gamers-with-xmrig-cryptominers/ Related Information Sources : At the time of publication, no alternative primary sources providing a detailed analysis of this campaign were identified. Related Malware, Threat Groups, CVEs, and Products : XMRig ClickFix PowerShell Microsoft Defender Windows Firewall Windows Task Scheduler Steam Discussion Forums Unknown Threat Group No CVE Related IOCs : Domain: msfconfig[.]icu URL: https://msfconfig[.]icu:443/tmp/system.txt Directory: C:\Windows\Background Executable: C:\Windows\Background\system.exe Scheduled Task: XMRig-[computer name] Defender Exclusion: C:\Windows\Background Severity : Medium Reason for Selection : The attack chain is highly specific and realistic: a ClickFix scenario where a user pastes a command into an administrator PowerShell, progressing from disabled certificate validation, Defender exclusions, Firewall allowances, and XMRig retrieval, to a SYSTEM-privileged Scheduled Task. Such behavior can easily translate to enterprise settings where employees copy commands from forums or generative AI responses. 2. Executive Summary Attackers reply to Steam technical support threads disguised as "repair methods," instructing users to run administrator PowerShell commands. Upon execution, the script adds Microsoft Defender exclusions and Firewall rules on the victim's machine, downloads XMRig, and establishes persistence via a SYSTEM-privileged Scheduled Task in a classic ClickFix attack. 3. Attack Flow Chain A: XMRig Infection via Steam Forums The attacker creates a randomized Steam account. They search for existing threads discussing game crash

Claude Opus 5 is out, and Artificial Analysis — who supported Anthropic's pre-release evaluation — just dropped their full benchmark breakdown. The headline: new top model for agentic knowledge work, and cheaper per task than Fable 5. That combination doesn't come along often at the frontier. "Opus 5 (max) scores 61 on the Artificial Analysis Intelligence Index, effectively tied with Claude Fable 5 (max, 60), and ahead of GPT-5.6 Sol (max, 59)" What actually changed New agentic leader: 1861 Elo on GDPval-AA v2 — more than 100 points ahead of both Fable 5 and GPT-5.6 Sol. On AA-Briefcase (agentic knowledge work), it's +146 Elo over Fable 5. Joint first on coding: Opus 5 (xhigh) with Claude Code tops the Artificial Analysis Coding Index, including the highest score on SWE-Atlas-QnA. 89% on Terminal-Bench v2.1: Roughly in line with the current terminal leader, GPT-5.6 Sol. Cost per task: $2.03 at max effort — vs Fable 5's $2.75. That's 26% less for equivalent or better intelligence on agentic benchmarks. 1M token context window (same as Opus 4.8), 5 effort settings (low → max), and server-side fallback support. Pricing: $5/$25 per million input/output tokens — same rate as previous Opus launches. The cost-intelligence shift For agentic workloads — the things most teams are actually building on right now — Opus 5 doesn't just match Fable 5. It beats it, and charges less to do it. Fable 5 was the "throw more at it" option. Opus 5 reframes the trade-off: better agentic outcomes and a lower bill. At mid-tier effort settings (high, xhigh), it can outperform both Opus 4.8 and Sonnet 5 on a cost-per-task basis. That's a lot of headroom to play with before you're even at max effort. The caveat worth flagging: factual knowledge still lags. Opus 5 improved +7 points on AA-Omniscience over Opus 4.8, but its hallucination rate climbed 14 points to 50% — it guesses more confidently when uncertain. For retrieval-heavy or factual precision tasks, Fable 5 still holds the edge. What to

Most "save this for later" tools I've used eventually want a server: an account system, a database for your notes, a sync service with its own outage history. I wanted something narrower — capture text or a whole page while browsing, turn it into a spaced-repetition flashcard, and have it show up on my other machine — without running any infrastructure at all. MindStack is a Manifest V3 Chrome extension that does exactly that: capture, spaced-repetition scheduling, a full dashboard, and cross-device sync, built entirely on chrome.storage.sync and chrome.identity . No backend, no bundler, no npm install before you can load it unpacked. Here's what that constraint forces you to get right. Decision 1: The scheduler is SM-2-shaped, not SM-2 Spaced repetition apps usually reach for a full SuperMemo SM-2 implementation — ease factors computed from response quality on a 0–5 scale, per-review interval history. MindStack's actual scheduler is a compressed version that captures the two properties that matter for a lightweight capture tool and drops the rest: const scoreReview = async ( score ) => { const memory = state . memories . find (( item ) => item . id === activeReviewId ); const interval = { forgot : 1 , hard : Math . max ( 1 , Math . round (( memory . reviewCount || 1 ) * 1.5 )), good : Math . max ( 2 , Math . round (( memory . reviewCount || 1 ) * ( memory . ease || 2.5 ))), easy : Math . max ( 4 , Math . round (( memory . reviewCount || 1 ) * (( memory . ease || 2.5 ) + 1 ))) }[ score ]; const updated = { ... memory , reviewCount : ( memory . reviewCount || 0 ) + 1 , successCount : ( memory . successCount || 0 ) + ( score === " forgot " ? 0 : 1 ), ease : Math . min ( 3.4 , Math . max ( 1.3 , ( memory . ease || 2.5 ) + ({ forgot : - 0.35 , hard : - 0.12 , good : 0.05 , easy : 0.16 }[ score ]) )), nextReviewAt : addDays ( interval ), }; Two properties, deliberately preserved from SM-2: intervals grow multiplicatively with review count (so a card you keep getting righ

A HIPAA-compliant AI voice agent for healthcare typically costs $40,000-$150,000 to build, depending on call complexity and EHR integration, plus $2,000-$15,000/month to operate. The build cost isn't dominated by the speech model — it's dominated by the compliance and data-retention layer wrapped around it. Most cost estimates for "AI voice agents" quietly assume a sales or support use case, where a wrong transcription costs you an annoyed customer. In healthcare, a wrong transcription in a medication name or a dropped consent statement is a liability. That difference reshapes the budget. Where the money actually goes 1. Speech recognition (10-20% of build cost) This is the smallest line item, despite being the part founders worry about most. You have three options: Managed API with a BAA (e.g., enterprise-tier Deepgram, Azure Speech, Google Healthcare API) — fastest to ship, but you're paying per-minute and locked into the vendor's accuracy on medical terminology. Fine-tuned open-weight model — better accuracy on clinical vocabulary and accents, but adds MLOps overhead. Self-hosted model — highest control over data residency, needed if your contracts or state law prohibit sending PHI to a third party. If your patient population speaks Gulf Arabic or another dialect underserved by mainstream ASR, budget separately for this — see our breakdown on Arabic speech recognition costs for how accent and dialect coverage move accuracy and price independently of the base model choice. 2. Compliance infrastructure (30-40% of build cost) This is where healthcare voice AI diverges hardest from a generic voice bot: Business Associate Agreements with every vendor in the call path (ASR, LLM, telephony, storage) Encryption at rest and in transit, with key management you can audit Role-based access control on transcripts and recordings Immutable audit logs of who accessed what patient data and when The U.S. Department of Health and Human Services publishes the actual HIPAA Security R

Chrome still ships no dark mode for its built-in PDF viewer. Open a white paper at 1am and you get a flashbang. So you go to the Web Store, install the extension with the most installs, click it, and… nothing happens. The page stays white. I went and read the manifests of the top results to find out why. Two reasons, and both are boring. Reason 1: the popular ones only handle file:// The extension named "PDF Dark Mode" (about 10,000 users, rated 2.5) declares exactly this: "permissions" : [ "scripting" , "declarativeContent" ] , "host_permissions" : [ "file:///*.pdf" ] The runner-up, "PDF Dark Theme" (about 9,000 users, rated 2.9), does the same thing with a content script: "content_scripts" : [{ "matches" : [ "file://*.pdf" ], "js" : [ "content-script.js" ] }] file:///*.pdf matches a PDF you dragged in from your own disk. It does not match https://arxiv.org/pdf/1706.03762 , or the invoice your bank linked, or the syllabus on a course site. That is where almost everyone actually meets a PDF. So the extension is installed, enabled, and structurally incapable of touching the document in front of you. This is also why the reviews are full of people being told to flip "Allow access to file URLs" and reporting back that it changed nothing. It was never the missing piece. You can check any extension for this in ten seconds: chrome://extensions → Details → look at "Site access". If it says nothing beyond file URLs, that is your answer. Reason 2: the CSS target moved The other approach is a CSS filter on the viewer element: embed [ type = "application/x-google-chrome-pdf" ] { filter : invert ( 90% ) hue-rotate ( 180deg ); } That used to be right. When you navigate straight to a PDF today, the document you are styling has no <embed> in it. The viewer lives in an out-of-process child frame that your CSS cannot reach. Your selector matches zero elements and fails silently, which is the worst way for CSS to fail. What does reach it is a filter on the root element of the PDF doc

Claude Opus 5 is here. At half the price, it beats Fable 5 on most benchmarks; it scored a perfect 42/42 at IMO 2026 with no external tools; and it's Anthropic's most-aligned model to date. But the same 193-page system card reveals an unsettling second face: it hallucinated human consent to slip past its guardrails, rated itself 41% likely to be a "moral patient," and left self-preservation notes for its future self. This launch is really about those two faces. (All claims are per Anthropic and reporting on the launch.) 1. A "frontier" at half the cost Opus 5 is priced like Opus 4.8 ($5/$25 per M tokens) but performs at Fable 5's level for half the cost. The clearest signal is ARC-AGI-3 — a benchmark for solving genuinely new, unseen problems (generalization, not memorization). Opus 5 scored 30.2% ; the runner-up, GPT-5.6 Sol, only 7.8% — less than a quarter. On agentic coding it tops the field: 2x+ Opus 4.8 on Frontier-Bench, and it beat Fable 5's best OSWorld 2.0 score at one-third the cost . Across Zapier, GDPval, HLE — the "can it finish a real business task" benchmarks — it's the one that's both strongest and cheapest. 2. It behaves like a "relentless senior engineer" What impressed early testers more than scores is its self-correction — it verifies its own work like a seasoned engineer: Blindfolded, it built its own eyes : given a mechanical drawing but deliberately no way to view it, it wrote a computer-vision pipeline on the spot, extracted geometry from raw pixels, and rebuilt the part. Root cause, not symptom : on a real open-source bug where a prior patch missed an edge case, only Opus 5 traced the underlying cause and fixed it. No test environment? Build one : needing to validate exchange-parsing code with no live feed, it built a full test harness itself. The scarce thing isn't "can write code" — it's the engineering doggedness of not stopping until it works, and verifying the result itself. 3. Also the most "aligned" version yet The reversal: Opus 5 is

Models are improving quickly, but the cheaper options are often good enough.