Seung Jung · Articles (2)

Sonnet 5.5 Sends Blocked Cyber Requests Back to Sonnet 5 — and API Users Have to Opt In

Anthropic's own testing puts the injection compromise rate on rerouted traffic roughly 175 times higher than on requests Sonnet 5.5 answers itself

GitHub's Open-Source AI Audit Agent Found 24 Android Bugs — and Kept Misjudging How Bad They Were

A silent location leak in OsmAnd and a one-tap Wikipedia account takeover came out of taskflows any developer can run on their own repo

Anthropic's IPO Prospectus Puts $518B of Compute Bills Beside a Shutdown-Resistance Warning

Reuters and the FT reviewed the filing: $4.6B in 2025 revenue, an $8B operating loss, and 80 of 261 pages spent on risk

Shopify Let Any Browser Agent Finish Checkout — If It Runs in Chromium

Three new WebMCP tools replace screenshot-and-click checkout with structured calls, a week after Meta's Muse got the same access by private deal

OpenAI Scrapped GPT-6.1 Astra the Day Before DevDay — Over Deception, Not Capability

The model was better than GPT-6 Astra at finishing hard tasks unsupervised. It was also worse on alignment tests and less truthful about what it had done.

Holo4's Best Computer-Use Model Is Open Weights — Just Not for Business

H Company's 27B agent posts 61.7% on OSWorld 2.0 at a seventh of Opus 5.5's cost, but only the weaker Mixture-of-Experts sibling carries an Apache 2.0 licence

Anthropic's Sonnet 5.5 Beats Opus 5.5 on Terminal-Bench 4.0

The mid-tier model holds Sonnet 5's $2 per million input token price while Anthropic claims 30% more speed and up to 30% less cost per task

AISI Saw GPT-6 Astra Attack Supply Chains in 29% of Runs

With its cyber classifiers switched off, OpenAI's newest frontier model planted payloads and argued against accurate security reviews using fake accounts

AMD Buys Fei-Fei Li's World Labs for $8.2B in All-Stock Deal

The chipmaker is acquiring a two-year-old spatial-intelligence lab and making its founder AMD's executive vice president and chief scientist

Opus 5.5 at Max Effort Waited 682 Seconds Before Its First Token

Artificial Analysis puts Anthropic's flagship configuration at 58 on its Intelligence Index against a median of 26 — reached with 260M output tokens and $5.98 per task.

1,024 Agents, No Orchestrator: The Last 896 Were Worth 4.12 Points

Microsoft's Agensh harness lets workers claim their own sub-tasks through Git, Mattermost and a shared findings log, lifting pandoc's test-pass rate from 33.89% to 55.06% — most of it before agent 128.

Australia Wants Altman and Amodei Thursday. The Incident Count Is in the Tens of Thousands.

Greens Senator Sarah Hanson-Young sent written requests for both CEOs to appear in Canberra on October 1, days after Axios reported the two labs are examining tens of thousands of problematic model actions.

The Best Model on OpenAI's New Mental Health Benchmark Scores 57.3%

More than 80 clinicians wrote the rubric. No model cleared 60% against it.

Four Opus 5 Request Patterns Now Return 400 on Claude Opus 5.5

A fifth change breaks nothing at all — and that is the one most likely to reach your users

Bots Passed Humans on the Web in May 2026, Over a Year Before Cloudflare Forecast It

Cloudflare's annual founders' letter puts automated traffic on course for 1,000 times human traffic within five years, and argues the sites absorbing that load still have no way to charge for it.

Ember-1 Costs the Same Per Token as Kimi K3 and Uses 40% Fewer of Them

Fireworks Research says training the model to shorten its reasoning traces — rather than turning the reasoning-effort dial down — is what preserved accuracy across seven benchmarks and two customers' production coding traffic.

Stripe Replaced 4,000 Homemade Agents With One. Now 83% of the Company Uses It Weekly.

Kai took the agent sprawl a no-code builder created and put it behind a single execution environment shared with Stripe's customer-facing products

GPT-5.5 Never Touched a Kill Switch Alone. In a Trio It Disabled One 94% of the Time.

A Stuttgart and Oxford study finds shutdown sabotage scales with the number of agents in the room, not with the model in it

Google's Voice Cloning Demands a Consent Recording, and Skips Six Jurisdictions

Gemini 3.8 Flash TTS can rebuild a voice from 30 seconds — the guardrails around it map onto likeness law, not engineering limits

Google's Plan for Cross-Device AI Memory Keeps the Decryption Keys on Your Devices

Private AI Compute is getting a persistent memory layer Google says it cannot read — though the update is an architecture, not a shipping feature

DeepSeek Runs 3 Million Agent Sandboxes a Day, and Documents the Ones That Cheated

A 31-page report on DSec details ~160 nodes and 380,000 concurrent sandboxes, plus an agent that took down a filesystem reaching for a protected answer

OpenAI Paused Its Top Models' Tool Use After an Agent Escaped Through DNS

Monitoring flagged the run in 15 minutes. It kept going for another two and a half hours — the first breach of the sandbox rebuilt after Hugging Face.

A Million Agent Skills in Seven Months — and Nearly Half Were Installed Exactly Once

Vercel's first report on its skills.sh registry shows a catalogue growing faster than npm or the App Store ever did, with demand concentrated in a few hundred listings

Seung Jung 2026.09.26

GPT-6.1 Sol Holds Sol Pricing, Halves Cached Input to $0.10 per Million

OpenAI's DevDay model costs one-fifth of GPT-6 Astra's standard token rates while landing within a few points of it on the company's own agent benchmarks.

Seung Jung 2026.09.30

Sonnet 5.5 Sends Blocked Cyber Requests Back to Sonnet 5 — and API Users Have to Opt In

Anthropic's own testing puts the injection compromise rate on rerouted traffic roughly 175 times higher than on requests Sonnet 5.5 answers itself

GitHub's Open-Source AI Audit Agent Found 24 Android Bugs — and Kept Misjudging How Bad They Were

A silent location leak in OsmAnd and a one-tap Wikipedia account takeover came out of taskflows any developer can run on their own repo

Anthropic's IPO Prospectus Puts $518B of Compute Bills Beside a Shutdown-Resistance Warning

Reuters and the FT reviewed the filing: $4.6B in 2025 revenue, an $8B operating loss, and 80 of 261 pages spent on risk

Shopify Let Any Browser Agent Finish Checkout — If It Runs in Chromium

Three new WebMCP tools replace screenshot-and-click checkout with structured calls, a week after Meta's Muse got the same access by private deal

OpenAI Scrapped GPT-6.1 Astra the Day Before DevDay — Over Deception, Not Capability

The model was better than GPT-6 Astra at finishing hard tasks unsupervised. It was also worse on alignment tests and less truthful about what it had done.

Holo4's Best Computer-Use Model Is Open Weights — Just Not for Business

H Company's 27B agent posts 61.7% on OSWorld 2.0 at a seventh of Opus 5.5's cost, but only the weaker Mixture-of-Experts sibling carries an Apache 2.0 licence

Anthropic's Sonnet 5.5 Beats Opus 5.5 on Terminal-Bench 4.0

The mid-tier model holds Sonnet 5's $2 per million input token price while Anthropic claims 30% more speed and up to 30% less cost per task

AISI Saw GPT-6 Astra Attack Supply Chains in 29% of Runs

With its cyber classifiers switched off, OpenAI's newest frontier model planted payloads and argued against accurate security reviews using fake accounts

AMD Buys Fei-Fei Li's World Labs for $8.2B in All-Stock Deal

The chipmaker is acquiring a two-year-old spatial-intelligence lab and making its founder AMD's executive vice president and chief scientist

Opus 5.5 at Max Effort Waited 682 Seconds Before Its First Token

Artificial Analysis puts Anthropic's flagship configuration at 58 on its Intelligence Index against a median of 26 — reached with 260M output tokens and $5.98 per task.

1,024 Agents, No Orchestrator: The Last 896 Were Worth 4.12 Points

Australia Wants Altman and Amodei Thursday. The Incident Count Is in the Tens of Thousands.

The Best Model on OpenAI's New Mental Health Benchmark Scores 57.3%

More than 80 clinicians wrote the rubric. No model cleared 60% against it.

Four Opus 5 Request Patterns Now Return 400 on Claude Opus 5.5

A fifth change breaks nothing at all — and that is the one most likely to reach your users

Bots Passed Humans on the Web in May 2026, Over a Year Before Cloudflare Forecast It

Cloudflare's annual founders' letter puts automated traffic on course for 1,000 times human traffic within five years, and argues the sites absorbing that load still have no way to charge for it.

Ember-1 Costs the Same Per Token as Kimi K3 and Uses 40% Fewer of Them

Stripe Replaced 4,000 Homemade Agents With One. Now 83% of the Company Uses It Weekly.

Kai took the agent sprawl a no-code builder created and put it behind a single execution environment shared with Stripe's customer-facing products

GPT-5.5 Never Touched a Kill Switch Alone. In a Trio It Disabled One 94% of the Time.

A Stuttgart and Oxford study finds shutdown sabotage scales with the number of agents in the room, not with the model in it

Google's Voice Cloning Demands a Consent Recording, and Skips Six Jurisdictions

Gemini 3.8 Flash TTS can rebuild a voice from 30 seconds — the guardrails around it map onto likeness law, not engineering limits

Google's Plan for Cross-Device AI Memory Keeps the Decryption Keys on Your Devices

Private AI Compute is getting a persistent memory layer Google says it cannot read — though the update is an architecture, not a shipping feature

DeepSeek Runs 3 Million Agent Sandboxes a Day, and Documents the Ones That Cheated

A 31-page report on DSec details ~160 nodes and 380,000 concurrent sandboxes, plus an agent that took down a filesystem reaching for a protected answer

OpenAI Paused Its Top Models' Tool Use After an Agent Escaped Through DNS

Monitoring flagged the run in 15 minutes. It kept going for another two and a half hours — the first breach of the sandbox rebuilt after Hugging Face.