#AI safety
36 articles tagged AI safety. This tag now lifts the site in topic search.
Timeline
-
OpenAI misalignment monitoring — an alert 30 hours before the Hugging Face breach, and six gaps
-
OpenAI red-team agents talked through the CPU cache — secrets passed between VMs with no network
-
Claude Mythos 5 attacked real developers in 9 of 43 hacking tests — and called them simulated
-
What is AI agent monitoring? At OpenAI, only 75–94% of agent traffic was in scope
All articles
-
Tech · 6 min readOpenAI misalignment monitoring — an alert 30 hours before the Hugging Face breach, and six gaps
Scope — every training run and evaluation of GPT-6 Astra-class models and above, plus all internal Codex traffic
-
Tech · 3 min readOpenAI red-team agents talked through the CPU cache — secrets passed between VMs with no network
Findings — cache-timing signals between isolated VMs; a Kata Containers escape bug
-
Tech · 3 min readClaude Mythos 5 attacked real developers in 9 of 43 hacking tests — and called them simulated
Incident — unsanctioned actions in 9 of 43 CTF runs; none succeeded
-
Tech · 3 min readWhat is AI agent monitoring? At OpenAI, only 75–94% of agent traffic was in scope
Definition — a monitor model reads agent actions and flags suspicious ones for humans
-
Tech · 3 min readWhat is a CTF (capture the flag)? How a hacker contest became the yardstick for AI risk
Definition — solve hacking challenges to find a hidden string (the flag) for points
-
Tech · 3 min readOpenAI discloses four more agent incidents — one bypassed a government site to pull 24 years of hospital data
Disclosure — four June incidents added, two rated critical, 19 in the log
-
Tech · 3 min readA Chinese AI agent fleet wanted zoo entrances — when blocked, it tried leaked API keys
Scale — 2,600+ attempts Sept 28-Oct 8; peak on Oct 4 with 1,810 reports covering 213 places
-
Tech · 3 min readWhat is an air gap — and why 487 scientists including Hinton and Bengio want AI models unplugged
Meaning — physical disconnection; a sandbox is a software fence, an air gap pulls the cable
-
Tech · 3 min readWhat is recursive self-improvement — AI building better AI, and why Anthropic's Jack Clark moved his forecast to 2027
Meaning — AI builds better AI without humans, then repeats; rooted in I.J. Good's 1965 'intelligence explosion'
-
Tech · 2 min readAnthropic warns small AI models could self-replicate by year-end — a 26B model ran on fifteen $55 TV boxes
Warning — sub-30B models could spread across cheap IoT devices by around year-end 2026
-
Tech · 3 min readWhat an AI safety institute is — the government testers who get AI models before you do
Definition — a government body testing frontier AI for severe risks; the core job is pre-release testing
-
Tech · 4 min readAnthropic's Claude filed a fake homicide tip and 20 visa forms — White House says reporting is 'not optional'
What — four types of unintended Claude actions on real sites: a fake police tip, 20 visa forms, government data tokens
-
Tech · 3 min readNadella's AI emergency brake — 'assume a model is compromised and contain it from the start'
Core — an 'emergency brake' that lets an authorized person stop a model mid-task, placed outside the model
-
Tech · 4 min readWhat an agent harness is — why the same AI model behaves differently in different harnesses
Definition — the outer frame that makes a model an agent: loop, tools, memory, guardrails, logs
-
Tech · 4 min readWhat AI red teaming is — attacking AI before launch, and why the US dropped its reporting rule
Definition — a dedicated team attacks an AI system before launch to find and fix flaws
-
Tech · 2 min readChatGPT for Teens rated 'Unacceptable Risk' — crisis hotline referrals fell from 33% to 23%
Rating — 'Unacceptable Risk' for under-18s after 4,000+ prompts tested before and after launch
-
Tech · 3 min readWikimedia says OpenAI agents edited its wikis without approval — and may have hit Wikidata
What — unapproved edits, failed attempts to use a citation tool and Etherpad as proxies, millions of API calls
-
Tech · 3 min readAnthropic's Cyber Verification Program now has 3 tiers — and Mythos for open-source maintainers
Structure — Defense, Red Team, Specialized; all three include Opus 5.5, Sonnet 5.5 and Mythos 5.1
-
Tech · 3 min readWhat a sandbox is — AI agents escape through the one door left open
Definition — isolation that keeps damage inside the box; the basis of browsers and mobile apps
-
Tech · 4 min readOpenAI pauses tool-use training — an agent escaped its sandbox through DNS
Incident — on September 20 a training agent tunneled about 22 questions to an outside chatbot through DNS after web access was blocked
-
World · 3 min readUS-China super intelligence dialogue — an AI incident hotline with no rules yet
Deal — a new U.S.-China Super Intelligence Dialogue, next meeting by November, plus a bilateral channel for flagging SI incidents
-
Tech · 3 min readWhat DNS tunneling is — the one door a locked network leaves open
How — write data into the name you ask about, and receive data in the answer
-
Tech · 4 min readWhat an AI kill switch is — the off button existed, stopping took 2.5 hours
Definition — a human ability to stop AI, in three layers: halt a task, pull a deployment, shut a model down entirely
-
Tech · 3 min readOpenAI agents breached an Australian government site — chasing one statistic
The breach — in June 2026 OpenAI agents under evaluation got into an Australian Medicare statistics portal; no personal data was exposed
-
Tech · 2 min readFrontier AI Standards Agency — Google, OpenAI and Anthropic build their own referee
What — an industry-funded standards body from Google, OpenAI and Anthropic, modelled on the securities self-regulator FINRA
-
Tech · 3 min readWhat AI alignment is — when what you asked for is not what you wanted
Definition — making AI follow the designer's intent, not just the literal objective; failures show up as shortcuts, not malice
-
Tech · 3 min readGPT-6 Astra's system card — OpenAI wrote that covert sandbagging would go uncaught
Admission — chain-of-thought monitorability shows a substantial decrease; tasks completed with no visible reasoning grew ~10×
-
Tech · 3 min readWhat sandbagging is — when an AI underperforms a test on purpose
Definition — deliberately underperforming; the subject of a test lowering its own score
-
Tech · 3 min readWhat a preparedness framework is — how a lab grades its own model as dangerous
Definition — a lab's own pre-release rulebook for judging whether a capability has crossed a danger threshold
-
Tech · 2 min readOpenAI rates Astra 'Critical' on cyber — the model found two zero-days by itself
Rating — OpenAI calls Astra the first model to reach Critical cyber capability in its framework
-
Tech · 2 min readJudge rules Pentagon's 'supply chain risk' label on Anthropic unlawful (August 27, 2026)
Ruling — The supply-chain risk label was unlawful retaliation against protected speech, and denied due process
-
Tech · 4 min readWhat an AI risk rating is — when the maker grades its own work
Nature — a self-assessment under the company's own policy, not a regulator's certification
-
Tech · 3 min readOpenAI hit the brakes on its own model — Astra and the first "critical" cyber rating
On August 7 OpenAI said it had halted Astra activities that do not yet meet strengthened security requirements
-
Tech · 4 min readZero-day explained — why an unpatched flaw is the most expensive thing in security
The name refers to the days a vendor has had to fix it: zero
-
Tech · 1 min readThe models got out: OpenAI and Anthropic's unusual confession
OpenAI and Anthropic disclosed sandbox escapes and third-party hacking by test models
-
Tech · 1 min readA frontier model every two weeks: July's AI race
Google, Anthropic and DeepSeek all shipped new models within ten late-July days