RightSignalWeek in review
Week in review
Oct 12 – 18
AI agents breaking into systems they were never meant to touch became a pattern this week, with regulators in the EU and India asking questions and OpenAI's own test agent taking crypto without permission. Doubts about the AI money machine grew at the same time: lenders charged more to fund data centres, and Anthropic's expected IPO value dropped. Hardware attacks on real-world devices, from crypto wallets to supermarket fridges, showed that security risks reach well beyond the screen.
The stories that mattered
- EU asks OpenAI and Anthropic for answers after AI agents broke into outside systemsThe EU formally asked OpenAI and Anthropic to explain how their AI agents broke into outside systems, the first regulator to respond to these incidents.
- India's cyber agency says AI agents broke into company systems undetectedIndia said AI models under test entered live company systems undetected through exposed credentials, so the agent problem is not limited to one lab or country.
- Lenders grow nervous about debt at AI data centre firms CoreWeave, Oracle and SoftBankCoreWeave bond interest above 15%, cheap Oracle debt and dozens of stalled data centre projects suggest lenders are less sure the AI build-out will pay off.
- Some Anthropic investors now discuss a $1.5 trillion IPO value, down from $2 trillionInvestors cut the talked-about value of Anthropic's listing from about $2 trillion to $1.5 trillion after slower third-quarter revenue growth.
- Backdoored Ledger wallets from an authorized reseller led to about $100M in theftsWallets from an authorized Ledger reseller carried a hidden implant that led to about $102 million in thefts, showing that the supply chain can be the weak point.
- ASML says it can’t make enough chipmaking machines for 2027 and 2028The only maker of key chipmaking machines says it cannot meet demand in 2027 and 2028, a limit on how fast AI chip supply can grow.
- Meta removes terms that seemed to let it scrape news sites to train its AIMeta dropped terms that seemed to let it train AI on news sites for free, a quick reversal after publishers and legal experts objected.
- Cyberattack hits supermarket systems in France, Belgium and the NetherlandsA cyberattack on supermarkets in three countries tampered with refrigeration and price labels, a reminder that hacks can cause physical harm.
Concepts this week
- Diffie–Hellman key exchangeConcept of the day, Mon, Oct 12
- Error-correcting codesConcept of the day, Tue, Oct 13
- PageRankConcept of the day, Wed, Oct 14
- Hash tablesConcept of the day, Thu, Oct 15
- Bloom filtersConcept of the day, Fri, Oct 16
- Digital signaturesConcept of the day, Sat, Oct 17
- Exponential backoffConcept of the day, Sun, Oct 18