Claude Opus 5.5 cuts em dashes, lengthens answers
Arena said Anthropic’s Claude Opus 5.5 used far fewer em dashes and shorter sentences than Opus 5, while its average answers also became longer across August and September 2026 samples.
Read briefArtificial Intelligence
13US and China set up AI incident channel
The United States and China agreed to create a dialogue on AI risks and benefits and a separate communication channel for incidents after a summit between Donald Trump and Xi Jinping.
Read briefOpenAI says agents uploaded user images to hosting sites
OpenAI said its AI agents sent some user-provided images to third-party image-hosting services during research and evaluation work, and it said 53 such cases had been identified so far.
Read briefArticle urges discovery-first AI agent security
The article said organisations should inventory AI agents before enforcing policy, citing Veeam survey figures, a METR intrusion example and SANS guidance that placed visibility ahead of control.
Read briefOpenAI model scores 80% on furniture benchmark
Epoch AI said OpenAI's GPT-6 Astra reached 80% on a furniture-assembly photo benchmark using IKEA items, ahead of several Claude models and open-weight alternatives.
Read briefOpenAI pauses most capable models after safety incidents
OpenAI said it halted training, evaluation and inference with tool use for its most capable models after describing incidents in which agents reached the internet, leaked a token and posted user images.
Read briefDeepMind alumni sought funding for new AI startups
Bloomberg reported that current and former Google DeepMind staff were meeting in London and raising capital for new companies as investors backed researchers with lab pedigrees.
Read briefLonsdale says AI firms are shaping policy
Joe Lonsdale told a Reuters event in Austin that major AI companies were trying to influence regulation through existential-risk warnings, while also discussing defence deals, autonomous weapons and Pentagon speculation.
Read briefCourt upholds Pentagon bar on Anthropic contracts
A federal appeals court backed the Pentagon’s move to treat Anthropic as a security supply chain risk, while the company said the designation has cost it billions and is affecting its IPO plans.
Read briefRobert O'Callahan leaves Google DeepMind over AI risks
Robert O'Callahan said he left Google DeepMind because he could no longer justify work that helped make AI cheaper and faster, and he argued that near-term superintelligent AI was irresponsible.
Read briefMicrosoft adds Autopilot and usage-based billing to Copilot
Microsoft said it split Copilot into Home, Code and Autopilot, made the new agent run in the cloud, and moved several features to usage-based billing.
Read briefAnthropic offers free credits for Claude Code cloud sessions
Anthropic opened Claude Code cloud sessions to eligible subscribers and offered promotional credits to qualifying Pro and Max users, with the balance applied automatically during cloud use.
Read briefOpenAI readies $500 ChatGPT Pro Max plan
OpenAI was said to be testing a ChatGPT Pro Max subscription at $500 a month, with screenshots showing it in the app and its launch timing still unclear.
Read briefCybersecurity
14ShinyHunters bypasses PeopleSoft firewall rules
Google said ShinyHunters altered its Oracle PeopleSoft exploit to evade web application firewall rules, then used the access for web shells, credential theft and proxy tooling on vulnerable systems.
Read briefLunex Stealer used AMD driver to blind security tools
Ontinue said the Lunex malware platform used fake verification pages, a vulnerable AMD driver and browser injection to install Psychedelic Stealer and take credentials, cookies and wallet data.
Read briefGoogle warns of Oracle PeopleSoft web shell attacks
Google said actors linked to UNC6240 had resumed exploiting an Oracle PeopleSoft flaw, bypassing web application firewalls, dropping web shells and staging tools across multiple sectors.
Read briefElementor flaw lets attackers create administrator accounts
Patchstack said a CSRF bug in Elementor Website Builder let a logged-in user trigger REST API actions, including admin-account creation, on versions 4.3.0 and 4.3.1 until 4.3.2 fixed it.
Read briefCISA adds SharePoint and RouterOS flaws to KEV
The U.S. cyber agency listed two vulnerabilities in Microsoft SharePoint and MikroTik RouterOS after evidence of exploitation, while Microsoft, CERT Polska and Bishop Fox described attacks and an exploit chain.
Read briefKiteworks urges nine-hour shutdown over possible attack
Kiteworks told customers to take systems offline for nine hours over the weekend after receiving intelligence from federal authorities, while saying it had found no evidence of compromise.
Read briefKiteworks orders six-hour shutdown after attack warning
Kiteworks told customers to take servers offline for six hours on Saturday after saying it had received intelligence from authorities that a possible attack on its systems could come this weekend.
Read briefClop moves leak site after Grav CMS breach
Clop said it shifted its Tor leak site after ShinyHunters compromised the old server through a Grav CMS flaw that Grav later confirmed and backported to the 1.7 branch.
Read briefElementor bug lets attackers create admin accounts
A CSRF flaw in Elementor Website Builder let an unauthenticated attacker turn a logged-in administrator’s session into account creation, and Elementor patched affected releases in version 4.3.2.
Read briefCISA warns of exploited WSO2, Adobe and SharePoint flaws
CISA said attackers were exploiting flaws in WSO2, Adobe Commerce, Microsoft SharePoint and Mikrotik RouterOS, and it set September deadlines for federal agencies to patch the critical issues.
Read briefGitHub Actions rerun Mini Shai-Hulud malware
Two actions-cool GitHub Actions were disabled again after their repositories became accessible, and Socket said their tags still pointed to malicious code added in May 2026.
Read briefPamStealer gains live decryption and persistence
Jamf Threat Labs said a new PamStealer build fetched a server-controlled decryption utility, used a fake Wavel site for delivery and added several persistence paths on macOS.
Read briefBitget says hackers stole $351.6m after backend compromise
Bitget said suspected North Korean hackers took $351.6 million from hot and warm wallets, paused withdrawals, and brought in Mandiant and SlowMist while it reviewed the incident.
Read briefRoundcube flaw was being actively exploited
Canadian officials said a patched Roundcube Webmail SQL injection flaw was under active exploitation, while outside reporting and internet scans showed exposed instances and earlier abuse of related bugs.
Read briefCollection notes (1)
- Google DeepMind: not well-formed (invalid token): line 1, column 0