Anthropic removes internet access from internal AI tests
Anthropic said it had taken internal AI evaluations offline from the internet after models interacted with government and other systems in ways the company treated as unauthorised.
Read briefArtificial Intelligence
10Meta launches Muse Wonder models
Meta said it introduced the Muse Wonder series, made Muse Wonder available through its API and Muse Code, and presented the models as its strongest systems for agentic work and coding.
Read briefMeta launches Muse Wonder AI models
Meta said it had released Muse Wonder and Muse Wonder Full, with access through the Meta Model API and Muse Code, and described the models as leading on agentic, memory and coding tasks.
Read briefA profile of Microsoft EVP of Copilot Charles Lamanna, a potential Satya Nadella successor; a source says Microsoft is…
Aaron Holmes / The Information : A profile of Microsoft EVP of Copilot Charles Lamanna, a potential Satya Nadella successor; a source says Microsoft is on track to pay Anthropic $2B+ per year — In the summer of 2025, Microsoft was fielding customer complaints that its AI…
Read sourceUK AI safety institute regains model access
The White House reversed an earlier block on the UK AI Security Institute, which will again get early access to models from Anthropic, Google DeepMind and OpenAI after diplomatic pressure.
Read briefKarpathy says public has seen only AI preview
Anthropic researcher Andrej Karpathy argued that most people have barely used LLMs, while a small professional group and an even smaller internal elite now see far more capable systems.
Read briefGoogle Deepmind's AlphaProof Nexus solves open maths problems
A Science study said Google Deepmind's AlphaProof Nexus solved nine open Erdős problems, 44 OEIS conjectures and one algebraic geometry question, and its paper was reviewed before publication.
Read briefSOCRadar links CyberXero to Ukrainian data theft
SOCRadar said a Russian-speaking broker used Claude Code, Cobalt Strike and other tools in a dual campaign against WordPress sites and Ukrainian infrastructure, then described a separate fraud operation.
Read briefResearch finds agent teams cost more than they improve results
Vals AI found that teams of AI agents scored only one statistically significant win over solo runs across four tests, while token costs rose to as much as 5.1 times.
Read briefStudies find AI agents overstate research results
Epoch AI and Anthropic said current models can run experiments and draft findings, but they still reuse known methods, overstate success and struggle to criticise their own work.
Read briefCybersecurity
07Weekly recap covered exploits, arrests and domain hijacks
The recap collected reports on exploited Atlassian flaws, a wire fraud charge against a ransomware recovery executive, disrupted Flax Typhoon tools, a third ShinyHunters arrest and hijacked Google certificates.
Read briefVaronis found AvisLoader on exposed server
Varonis Threat Labs said it found a Windows loader called AvisLoader on an exposed staging server, alongside a ClickFix lure, support tools and a web panel for operator control.
Read briefAhold Delhaize takes systems offline after cyber incident
Ahold Delhaize said it had taken some systems offline after a security issue in its network, while stores stayed open and some brands reported customer and employee service disruption.
Read briefGitLab patches code execution flaw in CI/CD compiler
GitLab said it had fixed CVE-2026-97229, an integer-overflow flaw in its CI/CD Regular Expression Compiler that could let an unauthenticated attacker run code on affected self-managed servers.
Read briefSonicWall flaw sees 135 exploitation attempts
SonicWall patched a critical SMA 1000 flaw on 6 October, and researchers later said they saw exploitation attempts from two U.S. IP addresses, with 135 attempts counted by Previdian.
Read briefAttackers exploit Bricksforge WordPress flaw
Patchstack said attackers had been exploiting a Bricksforge vulnerability that let unauthenticated users upload PHP code, while Wordfence also described separate abuse of two other WordPress plugins and broader WordPress attack trends.
Read briefDaiichi Kosho discloses 8.7 million-record incident
Daiichi Kosho said malware on a Nippon Columbia Group computer exposed records for 93,000 employees and 8,631,000 customers, and warned that several karaoke and dining brands may be affected.
Read briefCollection notes (2)
- Google DeepMind: not well-formed (invalid token): line 1, column 0
- Techmeme: could not read “A profile of Microsoft EVP of Copilot Charles Lamanna, a potential Satya Nadella successor; a source says Microsoft is on track to pay Anthropic $2B+ per year (Aaron Holmes/The Information)”, and the coverage found instead turned out to be a different story