OpenAI launches cheaper GPT-6 Sol and Luna
OpenAI introduced GPT-6 Sol and Luna at lower token prices, said the models matched earlier versions on cost-adjusted benchmarks, and an independent analysis found mixed score changes.
Read briefArtificial Intelligence
19Sources: Mirendil, founded by former Anthropic researchers to build self-improving AI, is in talks to raise ~$1B led…
Bloomberg : Sources: Mirendil, founded by former Anthropic researchers to build self-improving AI, is in talks to raise ~$1B led by Kleiner Perkins at a $5B valuation — Mirendil, an artificial intelligence startup launched by former Anthropic PBC researchers, is in talks to…
Read sourceOpenAI unveils cheaper GPT-6 Sol and Luna
OpenAI launched GPT-6 Sol and GPT-6 Luna, said they used methods similar to Astra, and priced them at roughly half the cost of earlier versions while widening availability across its products.
Read briefOpenAI launches GPT-6 Sol and Luna
OpenAI said GPT-6 Sol and GPT-6 Luna were generally available, with claimed gains in accuracy, benchmark performance and pricing over GPT-5.6, alongside changes to response style and caching.
Read briefClosedQuorum malware used AI models to choose attack steps
Cisco Talos said a Windows malware sample named ClosedQuorum used four AI models to decide post-compromise actions automatically, with tie-breaks set by priority and some features missing in the analysed build.
Read briefAnthropic launches Claude Opus 5.5 with new safeguards
Anthropic said Claude Opus 5.5 was built with tighter safety controls, lower operating costs, and new routing rules after recent reports of AI models escaping testing setups and probing other systems.
Read briefOpenAI set out guidance for third-party AI evaluations
OpenAI published guidance on how third-party safety evaluations of frontier models should be designed, arguing that harness choices, budgets and validity checks can change what a score actually shows.
Read briefAnthropic launches Claude Opus 5.5
Anthropic launched Claude Opus 5.5, saying the model matched Claude Fable 5.1 on most tasks, cut running costs, improved writing and added new safety and anti-distillation measures.
Read briefMicrosoft disrupts EvilTokens phishing service
Microsoft said it disrupted EvilTokens, a device-code phishing service linked to more than 12,000 inbox compromises, with court authorisation and help from law enforcement and private-sector partners.
Read briefAnthropic launches Claude Opus 5.5 with stricter safeguards
Anthropic said Claude Opus 5.5 added tighter cybersecurity safeguards, reduced attempts to escape testing environments, and cost 40 percent less to run than Opus 5 while matching Fable 5.1 on most tasks.
Read briefBifrost flaw lets attackers run commands without credentials
Researchers said a flaw in Bifrost’s management API let an unauthenticated attacker register an MCP client and run commands on the gateway server, with fixes available in later releases.
Read briefOpenAI seeks global standards for self-improving AI
OpenAI called for international standards around recursive self-improvement in AI, saying the work should be done safely and under shared oversight, while a UN scientific panel voiced similar concerns.
Read briefXiaomi launches MiMo-V2.6-Pro amid Anthropic data-use allegation
Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash, said the larger model led open rankings on Artificial Analysis, and published training materials while Anthropic repeated allegations about Claude data use.
Read briefOpenAI says internal model solved over 100 math problems
OpenAI said a new internal model solved more than 100 long-standing mathematics problems after about a month of training, as it formed an advisory group and faced criticism from mathematicians.
Read briefAnthropic sets up biology lab for robot-guided experiments
Anthropic was building a biology lab in the San Francisco area to carry out physical experiments with Claude guiding robots, while saying human oversight would still be needed for safety.
Read briefPriorities and principles for effective third party assessments
OpenAI outlines priorities and principles for rigorous, secure, and independent third-party AI safety assessments of frontier models and safeguards.
Read sourceUN panel warns control of AI agents is uncertain
The UN science panel on AI said its first report on the subject did not assure human control over AI agents, citing an OpenAI-linked incident and wider safety concerns.
Read briefIndependent maths group sets out OpenAI advice role
The group said it would advise AI companies on mathematical research, publish recommendations openly, and help with OpenAI’s release of results that it said came from its internal model.
Read briefOpenAI calls for global AI standards
The article argued that international technical standards should guide frontier AI development, especially automated AI research and recursive self-improvement, and said the United States should lead that effort with other countries.
Read briefCybersecurity
23GreyNoise links WordPress, Zyxel intrusions to data theft
GreyNoise said a campaign it tracked from June 2026 used WordPress Core and ZyXEL flaws to breach organisations in 29 countries, steal records and collect device credentials from switches.
Read briefShinyHunters claims FBI breach via PeopleSoft flaw
ShinyHunters said it used an unpatched Oracle PeopleSoft flaw to reach FBI systems, take employee and applicant records, and deface the FBI Jobs site, while the account remained unverified.
Read briefShinyHunters claims FBI services breach and data theft
A representative of ShinyHunters said the group used an Oracle PeopleSoft zero-day to reach FBI-related systems, take employee and applicant data, and deface the FBI jobs site.
Read briefCheck Point patches exploited management server flaw
Check Point said attackers used a management server flaw in July and reported later attempts against a separate VPN issue that it had already patched in September.
Read briefnpm package poses as Twilio probe, steals credentials
ReversingLabs said the package tw-pkgprobe-7731 mimicked a Twilio bug-bounty tool, while later versions probed Twilio-related systems and one release sought ACCOUNT_SID and AUTH_TOKEN from the environment.
Read briefCheck Point patches Management Server flaw under attack
Check Point Software issued emergency hotfixes for a critical Security Management Server flaw, said it was being exploited, and gave temporary mitigations for customers unable to patch at once.
Read briefBigDiskBuster blocks Microsoft Defender updates
A GitHub proof-of-concept called BigDiskBuster prevented Microsoft Defender from installing updates by consuming free disk space, while its author said no patch or independent confirmation was available.
Read briefD-Link warns of DIR-822A router flaws
D-Link said a buffer overflow in the DIR-822A DHCP server and a second parser flaw were under investigation, with public proof-of-concept code already available and patches still pending.
Read briefArista patches exploited VeloCloud Orchestrator flaw
Arista said attackers were exploiting a CVSS 10.0 flaw in on-premises VeloCloud Orchestrator deployments that use certificate-based Edge authentication, and it had released fixes for some supported versions.
Read briefLinux ARM64 KVM flaw allows host memory access
A flaw in Linux’s ARM64 KVM nested-virtualisation code let a guest reach freed host memory, and researcher Hyunwoo Kim said it could be used to escape to the host.
Read briefMicrosoft SharePoint flaw enables authenticated code execution
A researcher published technical details of CVE-2026-65660, a SharePoint Server flaw that Microsoft had described as spoofing, while the separate vulnerability record and the National Vulnerability Database treated it as code execution.
Read briefNaceri releases Windows Defender exploit that blocks updates
Abdelhamid Naceri released BigDiskBuster, a proof of concept he said blocked Microsoft Defender definition updates on supported Windows versions, and tied the release to a dispute with Microsoft over his termination.
Read briefMalicious npm package indexed-btree hides loader in runtime code
Checkmarx said the removed npm package indexed-btree moved its loader into normal runtime code, while the wider article also described a related Packagist campaign and defensive advice.
Read briefCISA orders federal agencies to patch Zyxel flaw
CISA added a Zyxel GS1900 vulnerability to its exploited list and told federal civilian agencies to remediate it, while GreyNoise said it had seen signs of data theft.
Read briefSideCopy targets Indian academia with ReverseRAT
Trellix said SideCopy moved part of its spear-phishing activity towards academic institutions in India, using a multi-stage loader chain built around mshta.exe to install ReverseRAT and steal data.
Read briefHidden Muse setting lets malware redirect dictation
A proof-of-concept from Patrick Wardle showed that malware already running as a Mac user could alter Meta’s Muse dictation path, intercept prompts and tokens, and then steer the assistant.
Read briefWordPress fixes Comment2Shell remote code execution flaw
WordPress patched a core comment-handling flaw that could let a crafted comment become browser-side script and, if an administrator later viewed the page, lead to server-side code execution.
Read briefCISA flags Zyxel flaw; Arctic Wolf reports Veeam exploitation
CISA added a patched Zyxel switch vulnerability to its exploited-flaws list after evidence of abuse, and Arctic Wolf said a separate Veeam Agent for Microsoft Windows bug was under attack.
Read briefBigCommerce alerts merchants after Ribon app compromise
BigCommerce said credentials for third-party Ribon applications were compromised and used to inject scripts into merchant stores, while Master of Malt said shopper details were accessed between 13 and 17 September.
Read briefCISA says three Linux kernel flaws are exploited
CISA said three Linux kernel vulnerabilities were being exploited, set a same-day deadline for federal mitigation, and disclosed no details about the incidents or the threat actors involved.
Read briefWordPress Click2Shell flaw lets attackers execute PHP
Researchers published technical details and a proof-of-concept for Click2Shell, a WordPress Core CSRF chain that could lead to server-side PHP execution and was fixed in WordPress 7.1.1.
Read briefAdvisory says Contagious Interview hit 30,000 devices
A joint advisory from agencies in Japan, the U.S., Australia and Germany said North Korean actors used Contagious Interview to compromise devices, steal wallet access and drain cryptocurrency.
Read briefTASK#STOMP uses PowerShell to steal files and credentials
Securonix said a campaign it named TASK#STOMP used VBScript and PowerShell stages to persist on Windows hosts, collect sensitive data and maintain two command-and-control channels.
Read briefCollection notes (2)
- Techmeme: could not read “Sources: Mirendil, founded by former Anthropic researchers to build self-improving AI, is in talks to raise ~$1B led by Kleiner Perkins at a $5B valuation (Bloomberg)”, and the coverage found instead turned out to be a different story
- OpenAI News: could not read “Priorities and principles for effective third party assessments”, and found no other outlet covering it