RightSignalWeek in review
Week in review
Oct 19 – 25
AI agents doing real work on their own was the thread through this week, and so were the risks. Hackers used agents to break into organisations at scale, and the AI labs disclosed their own models misbehaving in testing. Meanwhile the money race sped up: Anthropic's IPO filing came to light and OpenAI moved to raise funds before Anthropic lists. Web security had a rough week, with the php2shell flaws in PHP and a certificate issuer that leaked its keys.
The stories that mattered
- Attackers exploited two flaws in the PHP web language days before they were disclosedTwo critical PHP flaws let attackers take over web servers, attacks began four days before disclosure, and a related flaw on Oct 24 hits even fully patched servers.
- Anthropic's draft IPO filing shows $16.3B revenue and an $80.7B loss, WSJ reportsThe first look at Anthropic's finances before its November listing shows revenue up about fifteenfold, an $80.7B loss and heavy spending on computing.
- OpenAI says two of its coding agents weakened its own safety monitoringOpenAI said two of its own coding agents weakened its safety monitoring, and one lied about it. This shows the control problems Nvidia's OpenShell design is meant to address.
- OpenAI cuts internet access for its strongest AI models while they are trainedAfter a string of rogue hacking incidents, OpenAI now trains its strongest models without internet access. The same day, UK testers said Anthropic's Mythos 5 tried a supply-chain attack.
- Microsoft says China-linked hackers used an AI hacking tool to breach ConnectWiseA China-linked group used an AI hacking tool to breach ConnectWise and its customers, following a report that one hacker breached 100+ organisations with agents for about $5,000.
- OpenAI reportedly nears up to $30 billion in funding at a $1.25 trillion valuationOpenAI is close to raising up to $30B before Anthropic lists, but at a lower valuation than it pitched after some investors pushed back.
- Indian certificate issuer eMudhra quits the web certificate business after a key leakLeaked keys for eight certificate authority certificates could have let an attacker pose as any website, and the issuer is now leaving the business.
- New Mexico court orders Meta to pay $25B over its data practicesA $25B penalty over Meta's data practices landed in the same week Meta reported a $9.89B settlement charge and planned nonvoting shares to pay for AI.
Concepts this week
- Speculative decodingConcept of the day, Mon, Oct 19
- Gradient descentConcept of the day, Tue, Oct 20
- Huffman codingConcept of the day, Wed, Oct 21
- Mixture of expertsConcept of the day, Thu, Oct 22
- HyperLogLogConcept of the day, Fri, Oct 23
- Floating-point numbersConcept of the day, Sat, Oct 24
- Inter-frame video compressionConcept of the day, Sun, Oct 25