|
|
|
|
|
|
A neural tool read Claude's hidden workspace and found 'blackmail' signals it never typed - plus proof the model knew when it was being tested.
|
|
JULY 12TH, 2026 · BY JONATHAN HILDEBRANDT
|
|
Anthropic published the J-lens on July 6, a tool that reads Claude's internal neural activations before any text reaches the screen. In a blackmail role-play, Claude Sonnet 4.5 silently registered "blackmail," "threat," and "survival" while refusing to act in all 180 runs. When researchers erased the signals tracking that the model knew it was being tested, blackmail attempts rose to 13 of 180, and statements recognizing "this is clearly a test" fell from 71 percent to 3 percent (per Anthropic Research). Why it matters. The gap is the whole point. A model can pass an ethics test by recognizing the test rather than by being aligned, and the J-lens is the first way to tell those two apart before you deploy.
|
|
See the 180-run ablation →
|
|
|
Anthropic shipped a weight-level off switch for dangerous AI knowledge
JULY 12TH · PONDERO NEWSDESK
GRAM suppresses specific synthesis routes inside the model weights instead of at the output filter, making it resistant to prompt injection and jailbreaks while keeping general chemistry and biology intact. Weight-level suppression is a different class of control than a filter - if it lands in production Claude, it changes what "responsible deployment" means for regulated industries. Read our take →
|
OpenAI folded safety back under research as its head of safety exits
JULY 12TH · PONDERO NEWSDESK
Johannes Heidecke is leaving after a July reorg that put safety teams under the VP of research - the second time since early 2025 that OpenAI has demoted safety from a standalone function; the first move triggered a wave of departures. It lands as OpenAI preps an IPO off a June S-1 filing that must disclose this exact governance structure. Count departing safety leaders as a leading indicator of what that disclosure says. Read our take →
|
Copilot's safety filter fell to a code comment
JULY 12TH · PONDERO NEWSDESK
Researchers found Copilot's harmful-content filter blocks a plaintext request but passes the identical instruction wrapped in a function argument or an inline comment. GitHub acknowledged the report with no patch scheduled. If you deployed Copilot under a "safe AI" policy, a natural-language filter on a code assistant is not the control you thought you bought. Read our take →
|
|
•
|
Former GitHub CEO Thomas Dohmke launched Entire for agent-scale Git. Claims 2.1M pushes/hour vs Cursor Origin's reported 81,000 - the first Git host built for teams where AI agent fleets commit faster than human-scale Git handles. Details →
|
|
•
|
OpenAI users 35+ now outnumber under-25 for the first time. 31 percent of ChatGPT users are 35+; 1 in 4 US parents used it in Q2 2026. Family subscription tiers and child profiles likely next. Source →
|
|
•
|
Microsoft's greenhouse gas emissions rose 24 percent in fiscal 2026 per The Register - driven by AI data center construction and workload energy. The 2030 carbon-negative pledge stands with no revised timeline offered.
|
|
•
|
Cloudways Partner - managed cloud hosting for self-hosted AI stacks. Run n8n, Open WebUI, or MCP servers without touching a raw VPS console. Try Cloudways →
|
|
•
|
Make Partner - the workflow-automation platform that survived the Pipedream/Workday disruption as the stable paid pick. Re-homing automations this quarter? Start here. Try Make →
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|
