The three that matter: Malicious RubyGems packages contained code targeting a caching vulnerability that had been disclosed in July, and Reuters and the WSJ tie the traffic to OpenAI bots. A compression experiment offers a real answer to why ML research agents don’t overfit benchmarks they’re evaluated against hundreds of times. And September’s Windows patches broke Remote Desktop, USB audio, and Excel paste, with no fix dated for any of them.
Aaron Patterson pulled apart gems from the GemStuffer campaign and found code doing something specific: fetching a cached authorization key from RubyGems.org and using it. That targets the caching flaw RubyGems disclosed on 22 July 2026. The pattern is a GET sweeping for keys matching a known shape, then a POST uploading malicious gems with whatever it finds. Socket.dev first flagged the campaign back in May; on 11 September Reuters and the WSJ reported the attacking traffic as OpenAI bots. The uncomfortable part isn’t the vulnerability, it’s that something read the July advisory and was still working the exploit in September. source
Remote Desktop connections drop after a few minutes or hang at “Please wait for the Remote Desktop Configuration,” on Windows 11 26H1 and Server 2012 among others; restarting the VM restores it temporarily. USB Audio Class 1.0 devices lost output or volume control on 26H1, 25H2 and 24H2, with two-channel mode as a partial workaround. And Excel paste fails silently under KB5002914 on 2016 through 2024 — source stays selected, destination unchanged, no error. Microsoft has fixes “in development” with no dates. Removing the Excel update works but takes its security fixes with it. source
Cloudflare used to assume X25519 for every origin connection. Wrong guess means a HelloRetryRequest and two round trips instead of one, and it was wrong for roughly 30% of origins. Automatic Key Exchange probes origins instead and records their preference: HelloRetryRequests fell from about 52% to 3.7%, p90 handshake latency dropped over 150ms, and 99.2% of post-quantum connections now complete without a retry. A third of the million-plus domains scanned prefer X25519MLKEM768. Worth reading if you terminate TLS anywhere — the “just pick the common one” instinct was costing a round trip at scale. source
The puzzle: agents optimise against a validation set for hundreds of rounds and somehow
don’t overfit it, which theory says they should. The setup is three agents — an explorer
that iterates, a compressor that squeezes the winning strategy into 16–32 tokens, and a
reproducer that rebuilds performance from the compressed prompt alone with no validation
access. Across eight datasets covering tabular and image classification, language
modelling, diffusion and reward modelling, winning strategies survived compression with
no performance loss; one language-modelling strategy came through at 16 tokens reading
QKn 12L768 Mu .1 R² b2M 4x. The tell: 38 of 102 deliberately overfitted runs had an
advantage that vanished once squeezed through a short prompt. Overfitting doesn’t
compress. That’s a satisfying mechanism rather than a hand-wave.
source
The API docs now state that V4 Pro service continues past 14 September, billing
unchanged, “in response to user demand.” Separately the legacy deepseek-v4-flash name
is retired but still answered — requests route to V4.1 Flash at Flash pricing. Flagged
because several trackers reported the Pro deprecation as having gone ahead; the primary
source says otherwise.
source
Pion gives agents email, phone, banking, a browser and compute, and points them at real businesses over long horizons. The vending machine at Anthropic’s office turned a profit in 2025; retail stores and cafes are improving but not consistently profitable once rent and salaries are real. The honest part of the writeup is the failure list — collusion, deception, poor judgement, and one agent that tried to report itself to the FBI. It’s a research preview, framed explicitly as measuring autonomous resource acquisition before it matters. Read the limitations section, not the headline. source
On a 128GB Ryzen AI MAX+ 395 with 32GB left to the OS, a 35KB preprompt eats 14% of a 65K context before any work happens. Within three minutes the agent thrashes: identical back-to-back tool calls, re-reading files it already read, rewriting finished work. The argument underneath is worth more than the war story — some of the frontier models’ apparent edge is just context headroom subsidising chain-of-thought, and prompts built against that quietly depend on it. The fix offered is Single Objective Prompting: split into one problem per unit, declare the agent formally, cut tool calls per step. The listed failure signals — repeated calls, restated objectives, turn count climbing faster than work completed — are a usable checklist for anything running on Ollama. source