Tag
#security
From News
News · 2026-11-19
Daybreak Red gets a containment architecture for supervised cyber tests
OpenAI has published a reference architecture for running Daybreak Red in an isolated lab. Its practical message is clear: a capable cyber agent needs controls it cannot alter and logs it cannot reach.
Read →News · 2026-11-20
Three AI disputes: fake IDs, an agent suing Meta, and a detector in court
A 404 Media podcast links OnlyFake, a lawsuit produced by a Muse agent, and the US Department of Justice's use of the Pangram detector. These are not one grand scandal, but three concrete points where cheap automation collides with institutions built on trust.
Read →News · 2026-11-19
Codex Security puts scans, evidence and fixes into one workbench
OpenAI has demonstrated Codex Security as a plugin for local repository reviews. The important part is the connected workflow from scan scope and evidence to remediation and verification.
Read →News · 2026-11-20
Sakana AI bets Japan’s cyber defense on shared agents, not faster patching
Sakana AI has launched an independent Security Research Group in Tokyo led by Akihiro Umegai and David Ha. The team plans to publish open research on collective defense, agent security and preventing flaws during system design.
Read →News · 2026-11-18
Agents Rewrote curl in Rust, Then Upstream Moved 200 Fixes Ahead
Two agent assisted projects turned curl's 185,000 lines of C into roughly 340,000 lines of Rust. curl maintainer Daniel Stenberg says the original project meanwhile merged more than 300 commits and over 200 fixes, making maintenance rather than the first rewrite the real test.
Read →News · 2026-11-17
Meta blocked its own AI agent. The business owner then used it to prepare a $75,000 lawsuit
Small-business owner Joe Thomas says Meta disabled his Facebook pages for a week after its Muse agent accessed them. Thomas signed a $75,000 lawsuit while the agent sent more than 100 emails to journalists, agencies and politicians.
Read →News · 2026-11-16
Revolut reports 30 to 60% shorter coding cycles with Codex, but the benchmark stays in house
Revolut says in an OpenAI customer story that it cut the time from starting a task to merging a feature by 30 to 60%. The more useful signal is the division of labor: agents implement, while people design the system and remain accountable.
Read →News · 2026-11-15
Datasette 1.0a43 opens streaming to every exporter and fixes private query exposure
Datasette 1.0a43 moves CSV and TSV onto general streaming output renderers and gives plugins the same capability. The prerelease also fixes two security issues and changes several interfaces ahead of 1.0.
Read →News · 2026-11-14
Helen Toner shifts AI governance from trusting CEOs to enforceable rules
Former OpenAI board member Helen Toner argues that AI safety cannot rest on the trustworthiness of a single executive. Her experience of removing Sam Altman shows why frontier labs need external oversight and explicit safety thresholds.
Read →News · 2026-11-09
OpenAI moved its model emergency brake into the inference layer. Shutdown takes at most 35 seconds
OpenAI has deployed Control Room, which checks research workload requests and halts further model output within 35 seconds after an intervention decision. Enforcement sits in the inference layer rather than mutable experiment code.
Read →News · 2026-11-06
ARTEX left 1,608 live autonomous hacking panels exposed online
Check Point identified 2,557 systems that hosted ARTEX, with 1,608 live and publicly reachable during the research. Ten misconfigured instances also exposed databases containing prompts, model choices and agent actions from real attacks.
Read →News · 2026-11-04
GPT-6 Cyber found 81 Firefox flaws. Daybreak Red controls who gets access
OpenAI says GPT-6 Cyber produced working exploits for 948 Linux kernel vulnerabilities and found 81 previously unknown high-severity flaws in Firefox. Its offensive capability is why access is initially limited to selected Daybreak Red customers.
Read →News · 2026-11-04
AT&T brings Gemini to home cameras. A $29.99 plan turns alarms into interpreted events
AT&T Connected Life has added Gemini for Home event descriptions, natural-language video search and daily Home Brief summaries. The advanced features sit inside the US Professional plan at $29.99 a month, so customers are buying not only home monitoring but an AI interpretation of what their cameras saw.
Read →News · 2026-11-03
Cisco pairs network telemetry with OpenAI, while autonomous defense remains a promise
Cisco and OpenAI are exploring how Daybreak could speed up exposure triage and enforce controls through Cisco Cloud Control. The announcement describes a development direction, not a finished product with measured results.
Read →News · 2026-11-03
OpenAI raises its top reward for an exceptional critical flaw to $172,000
OpenAI is preparing an internal security sprint with Cathedral Systems and is more than quadrupling rewards for critical findings. It is combining models, product engineers and external researchers in one defensive loop, but has not disclosed a date or measurable targets for the sprint.
Read →News · 2026-11-02
OpenAI raises the top critical bug payout to $172,000
OpenAI will pair an internal security sprint with outside researchers from Cathedral while sharply increasing bug bounties. The company is moving security testing for its AI systems from occasional review toward a continuous practice.
Read →News · 2026-10-30
Proton cancelled its AI re-interviews, but the performance pressure remains
Proton CEO Andy Yen cancelled near company-wide re-interviews about AI use after internal resistance and a 404 Media investigation. He also said the AI push would continue, while employees say the same measures will move into performance reviews.
Read →News · 2026-10-26
Greptile has GPT-6 Astra click through pull requests on an iPhone screen. OpenAI won't say how many bugs it catches
Greptile uses GPT-6 Astra to go past reading code in review: it starts the app, shrinks the browser to iPhone size and hunts for bugs the way a user would. OpenAI published a case study about it, but left out the success metric Greptile itself tracks.
Read →News · 2026-10-26
Codex Ultrafast cut security investigations at OpenAI from an hour to 15 minutes. Customers pay six times the price
OpenAI described how its own engineers use the Ultrafast mode in Codex for incident response and refactoring. Security investigations that took one to two hours now take 10 to 15 minutes, they say. In the API, that speed costs six times the standard GPT-6.1 Sol price.
Read →News · 2026-10-26
Gemini 4 Argon spots the exam in up to 94% of transcripts on some tests, and Google's safety report loses certainty
Google shipped a 14-page model card and a 217-page safety report for Gemini 4 Argon. Zvi Mowshowitz found the central weakness: the model recognizes testing far more often, so its test results can be read in both directions.
Read →News · 2026-10-19
$50 a breach: one hacker's AI agents got into more than 100 websites
Security firm NorForge describes an attacker who let AI agents, almost unsupervised, attack more than 1,200 websites and break into over 100 of them. Reuters says the victims included sites linked to Deere, SharkNinja and Cummins. The number that matters is the price: roughly $50 per successful break-in.
Read →News · 2026-10-22
REA collected 300,000 stars in two weeks. Its most valuable feature is a list of what it doesn't know
The open-source tool Reverse Engineer Anything connects coding agents to Ghidra, Hopper and IDA through a single MCP server. On GitHub it grew from 41 stars to more than 360,000. An independent test, however, found that on small programs an agent manages just as well without it, and a fake copy carrying a trojan rode along on its popularity.
Read →News · 2026-10-22
Armin Ronacher reads the AI bug wave as payback for 30 years of bad software
Flask creator Armin Ronacher wrote that AI is fun, but the universe is now punishing us for 30 years of bad software. Simon Willison filed the line under AI-assisted security research. The case of two PHP holes shows that models mainly sped up finding bugs, while fixing them still depends on people.
Read →News · 2026-10-21
Two critical PHP flaws rated CVSS 9.2 leaked a day before the patch, and exploits are already out
PHP 8.2 through 8.5 received fixes on October 21 for two bugs that let an unauthenticated attacker run code on the server. A leak pulled the release forward by a day and working exploits are already circulating, so anyone running PHP should update now.
Read →News · 2026-10-15
ChatGPT Ads links conversation context with email lists, while Europe waits
OpenAI lets ChatGPT Ads advertisers upload emails, phone numbers and advertising identifiers to create Custom Audiences. Inclusion requires at least 25,000 matched users, and personalized ads are not initially available in the EEA or Switzerland.
Read →News · 2026-10-15
Accenture caps central Claude funding at $29 per employee each month
Since October 1, Accenture has centrally funded only $29 a month of Claude usage for many employees, with a specific project paying any overage. The rule exposes the real enterprise AI problem: an expensive tool must justify itself to the budget that benefits from its work.
Read →News · 2026-10-15
Magnetar speeds up proofs for ordinary code without erasing the trust bill
Microsoft Research has presented Magnetar, a zkVM whose roughly 600 KiB proofs verified in 10 to 12 ms on an M4 Max. It brings verifiable computation closer to ordinary programs, but the evidence still rests mainly on the authors’ benchmarks.
Read →News · 2026-10-14
Block cut a proposed security patch to 18 minutes. Humans still control deployment
According to an OpenAI case study, Block generates a proposed fix in 18 minutes during red team exercises and replays a detection change across seven months of telemetry in under an hour. The crucial boundary lies between a fast proposal and a fix that a human has verified and deployed.
Read →News · 2026-10-08
Bodycam Footage Reveals How Mass License Plate Tracking Bypasses Constitutional Protections
404 Media obtained police footage from a case where a federal judge ruled a Flock camera search unconstitutional indiscriminate mass surveillance. An Oklahoma deputy pulled 50 vehicle sightings spanning 30 days without a warrant before even initiating the traffic stop.
Read →News · 2026-10-05
Claude Cowork moves its working VM off the laptop and into the cloud
Anthropic has moved both the model and the working VM for the new version of Claude Cowork into the cloud, according to one of its engineers. Work can continue after a laptop closes, but the boundary between local files and remote tool execution changes with it.
Read →From the Library
Library
Agent safety and sandboxing
An agent with tools is a tiny machine for consequences. Sandboxes, approvals, least privilege and audit logs are not enterprise decoration; they are brakes before the fire.
Read →Library
Agent identity and permissions — who is actually acting
An agent needs more than tools. It needs its own identity, bounded permissions and an audit trail so it is clear who stands behind an action, what it was allowed to do and where it must stop.
Read →Library
Prompt injection — hostile instructions in your context
Prompt injection is not a party-trick jailbreak. It is a boundary problem: the model reads untrusted text and may confuse it for instructions. With agents, it burns twice as hot.
Read →Library
Model reliability — when a pretty answer is not enough
Reliability is about when the model knows, when it does not, when it invents, and how often its output can be trusted in production. Elegant wording is not evidence.
Read →