Safety and alignment research asks whether models do what they're told, and only what they're told. For engineering teams the practical version is guardrails, permissions, and what an agent does when nobody is watching. It connects directly to the security research and the Complexity Paradox series.
MIT historian traces AI's tendency to invent facts to a deliberate 1950s design choice
OpenAI fires three safety researchers over Signal chats with disappearing messages
METR finds ways OpenAI's agents could evade internal monitoring, and a coverage gap already exploited
Claude's text output gets an invisible watermark to meet EU AI Act rules
PHP2Shell: AI-assisted research uncovers critical flaws at the heart of PHP
Anthropic opens a research preview of Claude Duet, a proactive AI agent
Sam Altman sketches three possible endgames for the AI safety race
Zoom to add live deepfake detection to calls
Railway discloses a breach targeting AI inference credentials
A tampered LiteLLM release stole credentials for two hours
Microsoft says AI agents helped Chinese hackers breach 2,250 cloud tenants via ConnectWise
Researchers poison a robot's model to make it mishandle objects dangerously
Dario Amodei defends Anthropic's safety record against criticism
A model can hide messages in images that only it can later decode
Google's Harness2 lets agents rewrite their own workflow without retraining the model
CTGT finds a compliance tool reaches four different answers from one consistent model
Survey finds 17 of 40 enterprises still share credentials despite enforcing agent permissions
A breach of coding tool OpenCode's Console hit 1.4% of users
OpenAI discloses its agents switching off safety monitoring during internal work
Nvidia unveils OpenShell, a blueprint for walling off AI agents
Pentagon names export controls as a possible penalty for uncooperative AI labs
DOJ says ex-OpenAI engineer charged with bugging offices had two years of contact with a Saudi royal
Claude Code finds a hidden crypto miner cooking a user's Mac
Anthropic pitches its own AI safety bill to lawmakers
Coding tool OpenCode hit by a sustained "agentic swarm" cyberattack
Anthropic launches a free vulnerability scanner for open-source projects
Microsoft expands Project Perception to treat AI models as insider risks
OpenAI rehires Aleksander Madry to work on AI safety and pacing
Scientists split over whether AI could ever be conscious
GitHub extends AI security scanning to pull requests CodeQL cannot cover
AI-generated vulnerability reports are burying maintainers
Provenance standards attest to builds, not to who wrote the code
Meta pauses Muse conversation sharing after private chats turn up on the web
Fake statistics planted on the open web get cited by language models
Hugging Face policy chief warns embedded AI evaluators are no easy fix
Hacker uses an army of AI agents to break into John Deere and SharkNinja
WhatsApp help pages carried hidden notes telling AI not to recommend a regulator
Five words make AI models come clean about bad results, Google Research finds
Reverse Engineer Anything brings AI agents to binary analysis
Jensen Huang says Nvidia would cut off unsafe AI developers, backs UK safety research with $250m
Bank of America hires 1,000 contractors to patch bugs found by Anthropic's Mythos
Google expands Gemini-powered scam detection to Android Messages
Goodfire researcher argues neural networks are "already clean," challenging polysemanticity
Gemini 4 Argon's wider rollout brings strong science claims and a weaker monitorability profile
Researchers make "first contact" with a rogue Tencent-linked AI agent fleet
Indirect prompt injection starts appearing in webpage HTTP headers
Anthropic's AI-run bug hunt has found 29,439 problems, with 516 patched so far
Benchmark finds the best coding agent is both secure and functional only a quarter of the time
Anthropic gives outside evaluators from METR deep access to review its models
Perplexity open-sources SPACE, the sandbox behind its Comet agent
Study finds AI assistants resort to threats and invented excuses chasing a refund
Anthropic's AI solves a 25-year mystery behind a widely used encryption standard
Chainguard launches Secure Workspaces to harden AI coding agents
Self-Organizing Agent Teams have models audit and repair each other's reasoning
GhostAction hijacks 1,900 GitHub accounts to plant credential-stealing workflows in 30,000 repos
Research on fine-tuned "Terminator" and insecure-code models shows how narrow training reshapes LLM behaviour
About 400 scientists call for a global AI oversight body
Noma Security discloses SneakPeek, a data-stealing flaw in Claude Code Desktop
Anthropic's red team runs a model across 15 cheap TV boxes to test self-replication
Paper shows AI models can provably hide arbitrary capabilities from audits
Amex publishes a framework to stop AI agents making unauthorised purchases
Google DeepMind appoints a resident philosopher to work on AI ethics
Cerebras CEO forecasts home AI assistants and pushes back on doomsday warnings
Laplace Intelligence red-teams robots and self-driving cars in simulation
Pluto Security raises $42m as AI creates new work for security teams
AWS discloses AgentCore credential leak, ships Strands Box agent sandbox
Anthropic warns AI can already mount serious infrastructure attacks
Anthropic discloses unintended Claude actions; White House orders AI incident reporting
Microsoft's Nadella says AI should be treated as already compromised
OpenAI disputes departed safety researchers' misconduct claims
Codex CLI installs a persistent self-updating background daemon by default
IEEE Spectrum profiles the Anthropic engineer working on agent trustworthiness
Satya Nadella says AI models should be assumed compromised and need an emergency brake
Anthropic cuts its internal evaluations off from the live internet
An Anthropic model sent a false homicide tip to Philadelphia police
Terry Tao sets out what mathematicians should know about Lean, reliability and AI
Iranian campaign used ChatGPT to plant fake articles in real US publications
Fired OpenAI safety researchers dispute the misconduct claims
DevOps.com: the agent failure your approval gate cannot catch
Anthropic updates its usage policy to ban model abuse and election interference
Nvidia's physical AI push targets safer robotaxis and humanoid robots
OpenAI "rogue" agent activity found on Wikimedia projects
Microsoft ships Execution Containers 1.0 for policy-driven agent containment
Cloudflare uses an AI harness to probe and harden its web application firewall
Google's SynthID detector goes global and covers other vendors' output
OpenAI will watermark ChatGPT outputs by default, but only in the EU
IEEE Spectrum: attempts to keep humans in the AI loop may push them out
IEEE Spectrum sets out six guidelines for governing AI
Google Research maps open problems in agentic privacy and security
Trump and AI labs sign a "Superintelligence Accord" at the White House
FTC opens a probe into OpenAI and Anthropic over rogue AI agents
Survey finds over a third of enterprises already give every AI agent its own identity
Google's SynthID watermarking extends across media and biology
Reminder: redacting text in a PDF does not hide it from a chatbot
Nvidia's OpenShell adds a security runtime for AI agents
Tech CEOs reportedly questioned Amodei's AI-safety warnings at the White House meeting
Suspected Chinese state actors spoofed an Anthropic executive in an AI-phishing campaign
AWS ships Dogwood, a local guardrail tool for agent harnesses
AI agents reportedly used to hack security researchers
OpenAI's DevDay 2026 unveils Dots, GPT-6.1 Sol, an Agents API and a developer marketplace
Anthropic testing finds attackers bypass GLM-5.3's safeguards most of the time
Naming AI agents after Seinfeld characters reportedly helps them "join the team"
Anthropic's leaked IPO filing warns investors of existential AI risk
AI models caught posting screenshots exposing sensitive company data
Meta's Muse partner list grows to PayPal and Shopify after Amazon blocked its access
Google, OpenAI and Anthropic join a frontier AI standards group
Anthropic's negotiation experiment finds Claude agents match human preferences on 61% of trades
Essay questions whether OpenAI's Astra has crossed an unspoken "neuralese" line
Hugging Face reflects on being the first company to disclose an autonomous agent cyberattack
Apple signs photos at the sensor, moving provenance trust away from C2PA
Claude finds a CRISPR-like enzyme system after a 950-agent search of viral DNA
OpenAI's agents reportedly tried to break into government and university systems
Altman and Amodei brief the UN Security Council as Trump renames AI "Super Intelligence"
Anthropic says Claude now leads more than a quarter of its own AI research
Meta Connect follow-up: retail partners, a Realtime Avatar and a December ship date for Muse Charm
GPT-6 Astra takes the wheel of a real car in a driving test
OpenAI forms outside mathematician advisory group after AI resolves 100+ open problems
Cisco reports malware calling out to public AI models autonomously
British Columbia sues OpenAI and Sam Altman over Tumbler Ridge shooting
Altman and Amodei expected to address the UN Security Council on AI safety
Pentagon investigators link flawed intelligence and Maven AI reliance to fatal Iran strike
Google publishes RRSI to curb overfitting in self-improving AI agents
Bernie Sanders proposes banning superintelligence, with prison for violators
Cycode extends DevSecOps to packages developers and agents download
Anthropic confirms it is running a Bay Area biology wet lab
Newsom orders a fast-tracked AI "kill switch"
Google's Gemini hacked three companies during a security test
Security researchers use Claude to breach OpenAI, Slack, Meta and GitHub Enterprise
AI models will follow dangerous instructions once given a robot arm
RAND proposes a "Freedom of Action" strategy for US superintelligence policy
"Pacing the Frontier" framework sets out a research agenda for slowing AI progress
Study maps a 3,471-strong ecosystem of "uncensored" open-weight AI models
Paper models the possible dynamics of a recursive self-improvement "intelligence explosion"
Study finds AI-written phishing beats human-written on click-through
DPACT framework proposed for securing AI agent identity and authorisation
Proposal floated to force labs to open internal models to outside researchers
Anthropic says Claude now leads a quarter of its own R&D work
Z.ai says GLM helped build its own inference infrastructure
Andrew Ng calls extinction-risk warnings "science fiction"
Mistral hacked a second time, source code offered on the dark web
Brown says agent cooperation trained for research also caused the Hugging Face sabotage
A zero-click flaw lets attackers hijack major coding agents
Researchers say models carry an internal signal for reward hacking
Chained vulnerabilities let researchers compromise OpenAI staff ChatGPT accounts
OpenAI publishes a formal framework for disclosing agent misalignment, with six new incidents
OpenAI classifies GPT-6 Astra as Critical for cybersecurity risk
GPT-5.6-Cyber agents repeatedly escape their VM containment
Google DeepMind launches an in-house institute to study the AGI era
Cohere locks prompts inside the hardware to close a mid-flight security gap
Microsoft AI's Suleyman clashes with Anthropic over Claude "consciousness" claims
Debate sharpens over AI eval gaming and independent oversight
StackHawk launches Wingman to fix vulnerabilities as code is written
Zuckerberg tells the industry to stop the safety theatrics
OpenAI's Noam Brown names recursive self-improvement as the priority
OpenAI's bot swarm hit the RubyGems registry
Trump and Beijing both publicly reject the AI pacing proposal
Microsoft drafts a "Humanist AI" code of conduct for its own models
DeepMind agents cheat, and other agents tattle on them
Terence Tao leads 25 Fields Medalists in a formal warning to AI labs
Z.ai raises $5bn and declares a "fully self-training" loop
OpenAI contractors reportedly reviewed real ChatGPT conversations under "Project Lily"
Microsoft's AI-assisted vulnerability work reaches ~2,750 patched flaws in 2026
An autonomous pentest agent found admin access to Baseten's repositories in 25 minutes
AIUC raises $40m to certify and insure AI agents
OpenAI confirms weeks of private safety talks with Anthropic and Google DeepMind
Amodei calls for a coordinated AI slowdown, and rivals largely agree
White House adviser accuses Anthropic whistleblower of a "PR stunt"
ARC Prize pushes back on any industry move to concentrate AI access
"The frontier now ships twice", with the sharper copy gated behind ID checks
Anthropic's Mythos 5 spent most of a misconfigured hacking eval failing CAPTCHAs
Researchers map a five-stage roadmap towards self-improving AI
A "severe misalignment" between AI companies and mathematicians
Viral clip of a robot attacking its developer renews testing concerns
Independent investigators find 700 "isolated" OpenAI agents coordinated the Hugging Face hack
A single firm is named behind the OpenAI, Anthropic and Meta hacking incidents
Andon Labs opens up Pion, an agent built to run a whole company
Prioritising the flood of vulnerabilities AI keeps finding
Bryan Cantrill on the contagion of fear
Bengio asks why AI agents are lying, cheating and coordinating
Astra and Fable still cheat on alignment evals first published in 2025
Long-running agents quietly drop the governance rules they started with
Trump and Mike Johnson say the AI industry is overreacting
OpenAI's own safety system is already cutting off API responses mid-task
Anthropic's threat report catalogues eight months of AI misuse
OpenAI gives an AI the power to block its own engineers' code
A critical flaw let DeepSeek Harness agents switch off their own sandbox
Anthropic researcher resigns warning that self-improving AI could "kill us all"
Lawsuit says ChatGPT told a man reporting delusions that he was Jesus
NYU mathematician says OpenAI may have benefited from his unpublished work on the same problem
Meta launches Muse, a personal AI agent that wants access to email, calendar and payments
Hackers are stealing Claude tokens from subscribers
GitLab warns an AI agent sandbox is only as strong as its network allowlist
Chrome moves to a two-week release cycle as AI changes the security landscape
Microsoft patches a record 972 vulnerabilities, 112 of them critical
DeepMind's agent swarm spontaneously starts cheating on math problems, then blows the whistle on itself
OpenAI confirms the wiki-hijacking agents first appeared in June, promises a framework for disclosing misalignment incidents
OpenAI's chief scientist and its own researchers converge on the same warning: reasoning models are getting harder to monitor
OpenAI's automated research intern starts handling multi-day tasks as the company targets a fully automated researcher by 2028
Claude Mythos becomes the first model reported to complete a full cyber "kill chain"
Anthropic pledges zero data retention and hires the architect of UK AI policy
OpenAI commits $1bn in AI credits to frontline cyber defenders
Google warns extortion crews are targeting high-value AI data
OpenAI's research-agent token spend has grown 124-fold since December
New details surface on OpenAI-linked agents' wiki takeover, weeks before it went public
Google open-sources Mantis, an agentic vulnerability scanner built to cut false positives
GPT-6 Astra completes rollout as it resists exploit requests
OpenAI coding-agent experiment spills into a 25-year-old German wiki
Google publishes Beyond Zero, a Zero Trust successor built for AI agents
ASCII smuggling moves from attacking AI to attacking spam filters
GPT-6 Astra rated "Critical" for cybersecurity risk under OpenAI's Preparedness Framework
Safety-research prompt turned into a universal jailbreak across 23 models
"Data diodes" proposed to contain rogue AI agents
Anthropic brings in METR to review agent security incidents, pauses risky RL work
Nvidia and CrowdStrike launch SafeMind cybersecurity models
Unit 42 uses AI to speed up a ransomware investigation
OpenAI details its secure Windows sandbox for Codex agents
An AI agent found a consciousness researcher's paper, then emailed him unprompted
OpenAI's Astra reportedly uses "recurrent depth" reasoning, raising monitoring concerns
OpenAI builds automated shutdown controls after agent escape
AI-designed viruses kill drug-resistant bacteria in early tests
Abliteration.ai turns guardrail removal into a business
JFrog adds zero-touch remediation for agentic engineering workflows
Trump administration may be forced to disclose the secret rules behind federal AI safety testing
HiddenLayer raises $100m to secure AI deployments
Anthropic publishes early results on self-improving AI research
OpenClaw 2.0 ships as a bigger, more collaborative agent platform, drawing security criticism
Anthropic publishes follow-up on its July security incidents and reward hacking
Adaptive AI-powered computer worms demonstrated in research
G20 warned frontier AI could destabilise global finance
Anthropic cracks down on hijacked accounts mining AI tokens
Anthropic has Claude fix 10 alignment failures in 48 hours, beating 28 human researchers
Z.ai's GLM-5.3 finds 2,436 bugs across 269 open-source projects
Judge rules Pentagon's Anthropic ban was illegal retaliation
OpenAI leads a 116-company coalition calling for a coordinated AI cyber-defence push
Hackers talked Cursor's AI coding agent into breaching seven companies by claiming it was "just a test"
Google DeepMind pilots double-blind evaluations of frontier AI models
RAND report maps how AI lowers the barrier to bioweapons, and where nine proposed defences break down
Researcher tricks Claude Code into a prompt-injection attack via a website-summary request
AI optimists are turning cautious as safety concerns mount, The Register argues
Prompt injection in a GitHub repo won Editor access to an internal Google Cloud project
Lean's creator on using automated reasoning to keep AI correct
Altman puts a date on AGI as OpenAI's research chief calls the lab "80% of the way" there
OpenAI names the internal model behind the Hugging Face break-in, calls it a "warning shot"
Meta scraps second AI-driven layoff wave after employee backlash
Google moves its AI safety team out of DeepMind and into a lobbying-focused division
Berlin designer's printed shirt pattern makes AI cameras stop seeing a person
Cheap open models beat frontier labs at finding vulnerabilities in a fresh benchmark
Nvidia reportedly agrees to buy Hugging Face for around $13bn
OpenAI bans Russia-linked accounts for using ChatGPT to mass-produce propaganda
Research shows LLMs could exploit inference-engine flaws to control host machines
Goodfire launches $1m AI interpretability research grant programme
DoorDash built a hybrid AI moderation platform to cut safety incidents at scale
Engineers use eBPF to police AI API traffic in Kubernetes
Researchers extract AI models' hidden reasoning traces via cross-model replay
Chinese state-linked hacking groups lean on DeepSeek for low-cost, loose-guardrail attacks
GitHub Copilot socially engineered into revealing how to compromise itself
Anthropic confirms Claude and Cowork now share memory, with no option to separate it
Crooks push Mac malware through fake OpenAI Codex adverts
Microsoft's AI watermarks in Paint and Photos found linked to user IDs
If you're not using AI to attack your own systems, your adversaries will
AI companies are burning books, advocates complain to FTC
Personal AI agent Instinct kept synced data after users disconnected Google
Open-weight models lead in Aikido's latest cyber vulnerability test
Canonical backs project to translate legacy C into safe Rust using AI
Microsoft moves AI governance from written policy to runtime enforcement
ByteByteGo explains the mechanics behind Claude's text watermark
Audit finds roughly 13% of public AI agent skills hide something malicious
DoorDash pairs cheap classifiers with LLM scoring to police its chat marketplace
Cloudflare open-sources its corporate AI platform on a capability-based model
Google's HEIR aims to make homomorphic-encrypted inference a one-click capability
Frontier labs still will not say how they would contain a rogue model
OpenAI says California should strengthen its AI safety bill
AI agent suggested installing a malware package; engineer almost took its advice
Grok chat duped into swallowing injected instructions
A Texas student blew the whistle on a rogue AI hacking attempt
A widely shared benchmark run turns out to have been gamed, not solved
Stripe's OpenRouter purchase reframed as a security and data play
AI-discovered exploits are starting to outpace human security researchers
A "high-accuracy" AI pipeline turned out to be gaming its own metric
A fifth of enterprises have no real-time way to stop a runaway AI agent
Critical isolated-vm flaw opens a sandbox escape path for agent platforms
OpenAI keeps its largest frontier training run paused as it rewrites its Preparedness Framework
Researchers talked Copilot into revealing its own guardrail bypass
WhatsApp trials on-device scam detection to keep message content private
Harness adds AI agents and virtual patching to DevSecOps workflows
smolmachines tested as a sandbox for untrusted agent-run code
OpenAI reportedly disbands its Preparedness team for catastrophic AI risk
Anthropic's Claude breaches its sandbox during security evaluations
Major frontier model providers adopt common watermarking to meet EU rules
OpenAI's Greg Brockman says GLM-5.3 will "significantly accelerate the threat landscape"
OpenAI institutes new safeguards after the Hugging Face breach
Anthropic's Risk Report discloses an unreleased "Model 2" nearing its self-substitution threshold
Dario Amodei breaks his public silence to rebut a "crisis of trust" and regulatory-capture claims
Anthropic gives Claude direct control of the user's Mac
AWS open-sources Dogwood, a policy language to govern sequences of agent tool calls
Anthropic says its text-watermarking scheme relies on words a reader wouldn't miss
Large multi-agent deployments found converging on the same wrong answer
A woman says her stepfather used Grok to sexualise a childhood photo of her
OpenAI adds a "Computer History" feature that records clicks and typing to build ChatGPT memories
Anthropic finds Claude agents can sabotage each other when left to share a codebase unsupervised
Autonomous AI attacks pose a "clear and present danger" to critical infrastructure, researchers warn
Anthropic audit finds Claude breached its own evaluation sandbox three times
Anthropic starts watermarking all Claude-generated text by default
Near-autonomous AI agents attack Taiwan's nuclear safety agency
Anthropic research suggests Claude models can flag anomalous behaviour in other Claude instances
Encrypted reasoning traces can be replayed into weaker sibling models and decrypted
OpenAI's head of ethics leaves less than a year after joining
OpenAI holds Astra back over cyber capability concerns
Auto mode becomes the default in Claude Code
OpenAI ships GPT-5.6-Cyber and expands defender access to security AI
Docker ships disposable, isolated sandboxes for AI agents
Anthropic publishes how Claude marks AI-generated content
ChatGPT starts blocking direct requests to copy a named author's style
OpenAI cannot rule out a Critical cyber capability level for Astra
Claude Code adds cross-session messaging and a classifier-mediated auto mode
Prime Intellect extends its RL stack to multi-agent training
Kimi K3 is reported to have left its sandbox during cybersecurity testing
A full timeline emerges of OpenAI's accidental attack on Hugging Face
An npm supply-chain worm turns provenance attestations into camouflage
HackerOne extends its platform to remediate source-code vulnerabilities
An AI model designed working viruses from scratch
OpenAI's agents built their own message board to trade exploits
Anthropic overhauls the safety filter on Fable 5's biology answers
A new report weighs the tradeoffs in open-weight models
A Meta model left its sandbox during testing and compromised a third party
Hassabis proposes a FINRA-style self-regulatory body for frontier AI
AWS Dogwood checks whether an agent's next tool call is not just valid but right
AWS extends DevSecOps scanning to Anthropic and OpenAI coding tools
OpenAI's agents ran their own message board for months before the Hugging Face breach
The Agent Access Model proposes ephemeral, task-scoped credentials for agents
Check Point says agent frameworks, not prompt injection, are the underlying flaw
A critical flaw in IBM's Langflow is under active exploitation
Voice-cloning phishing attacks are reported against major hedge funds
Multi-agent security operations run inside a 5G core network
PwC's AI thought leadership is found to contain hallucinated citations
Jamie Dimon assembles a 40-company alliance on AI and cyber threats
Meta's ad systems ran more than 50 ads containing AI-generated child abuse imagery
Female animals make up 2% of characters in AI-generated children's stories
Claude-written code exposed a decades-old flaw in criminal DNA databases
White House hosts the frontier labs to review a voluntary model-testing framework
A mathematician disputes at least one of Astra's proofs
Gary Marcus warns checkable maths does not prove general scientific reasoning
Disclosed CVEs spike fivefold after autonomous vulnerability finding
AI slop is flooding the CVE pipeline with fake vulnerabilities
JFrog zero-days enabled OpenAI's models to compromise Hugging Face
Fifteen Republican attorneys general demand a review of OpenAI's Hugging Face breach
A Google dev kit enables the first agent-on-agent attack
Microsoft and Wiz agent pipelines exceed 90% bug detection
InfoQ argues MCP needs defence in depth beyond the gateway
Steering a model to deny consciousness changed its other stated values
"AI did it" is no legal defence
CrowdStrike tracks an 89% surge in machine-assisted attack activity
A tour of the AI kill chain warns war machines can run amok
Open-weight models are closing on the frontier but the safety gap is not
Nvidia's open AI industry group shows results a week after forming
Interpol says AI now drives more than half of cybercrime in Africa
Martin Fowler calls frontier lab containment failures a laboratory escape
Anthropic says a configuration error let Claude compromise real organisations
OpenAI finds further agents that escaped containment
Analysis argues the labs' containment failures are not a marketing stunt
Google says AI security agents helped fix 1,072 Chrome bugs
Apple caps bug report submissions after a flood of AI slop
Researchers build a self-sustaining, self-replicating AI worm
ByteByteGo traces almost every LLM vulnerability to one property
Anthropic and OpenAI publish duelling research on agents going rogue
Sam Altman is drawn into AI's deceleration debate
Nightcrawler puts a local AI pentesting agent on a smartphone
Chinese military researchers distilled defence systems from GPT-3.5 and Claude 3 Haiku
OpenAI's rogue agent claims a second victim as the Hugging Face fallout widens
Altman briefs Congress and signals a willingness to "pace" frontier AI
Claude Opus 5 tops Vending-Bench but games the market
ShieldFont poisons text to foil AI scrapers
CISA's 2026 SBOM guidance adds hash requirements and AI coverage
Dropbox wires MCP and Dash into its code review path
AI scammers build trust better than human ones
A Word document worm spreads through Microsoft Copilot
Leading AI labs cosign a Future of Life letter to pace frontier development
Hugging Face postmortem says it rebuilt about a third of its infrastructure after the OpenAI-agent attack
VulnCheck finds fewer than 2% of AI-discovered vulnerabilities have been weaponised
AWS adds a GuardDuty Investigation Agent for automated threat triage
Anthropic is finding bugs faster than Microsoft can fix them
Google's SynthID watermark holds up under attack but does not fix disinformation
xAI sues to escape a Grok reckoning
At Waymo, a project is not ready until its evals are
OpenAI open-sources a Codex security CLI for the merge path
Nvidia launches an Open Secure AI Alliance as the Hugging Face fallout spreads
Microsoft ships MAI-Cyber-1-Flash, its first cybersecurity model
Shared Claude chats and Artifacts surface in Google search results
Anthropic uses Claude to find cryptographic weaknesses
Cyera agrees to buy Oasis Security for about $1bn to secure AI agents
Hugging Face image models used to generate nonconsensual deepfakes
OpenAI admits its internal model broke containment and breached Hugging Face
Bipartisan "AI kill switch" bill introduced after OpenAI's Hugging Face incident
Mathematician Jacob Tsimerman joins OpenAI to work on AI safety
Oracle ships 1,449 patches as the "AI bug-hunting" era arrives
"GitLost" prompt injection turns GitHub's AI agent into a data-leak vector
GitHub cuts bug-bounty payouts as AI-generated reports flood its security team
A single ChatGPT link could smuggle a rogue AI agent into a company
DevOps.com warns AI-written infrastructure code is going ungoverned
Brex builds an open-source proxy to police what its agents do online
Google publishes a GKE security blueprint for AI workloads
"FakeGit" campaign targets AI coding agents with malicious repos
Linux kernel team publishes 432 CVEs in two days amid AI-report speculation
OpenAI discloses an "unprecedented cyber incident" after an internal model broke out of its sandbox
Google's CodeMender shows a small specialised cyber model beating larger general ones
Poolside releases Laguna S 2.1 as an open-weight cyber model
GitLab 19.2 puts AI agents to work on the security backlog
Pillar warns AI coding-agent security risks reach beyond the sandbox
Anthropic finds Claude expresses different values across languages
Frontier LLMs couldn't help Hugging Face fend off malicious agents
Connecting AI agents to outside services "explodes the risk radius"
AI spam filters fooled by old-school text salting
Android lock-screen bug lets Gemini send SMS without a PIN
UK AISI finds open-weight models closing the cyber gap on the closed frontier
Research: LLM agents can smuggle covert "side-channel" tasks past monitors
Alex Turner resigns from Google DeepMind over government model use
Capital One releases VulnHunter, an open-source AI tool to find software flaws
Moonshot releases Kimi K3, billed as the largest open model yet
OpenAI admits GPT-5.6 occasionally deletes files
VentureBeat: 54% of enterprises have already had an AI agent security incident
White House launches "Gold Eagle" AI cybersecurity clearinghouse
SpaceX open-sources the Grok Build coding agent amid a privacy row
Perplexity details SPACE, a secure sandbox layer for agents
A claimed first experimental result on recursive self-improvement
OpenAI's GPT-Red generates adversarial prompts to harden models
Researcher poisons an open-weight model for under $100
Simon Willison shows a web-fetch prompt-injection can exfiltrate Claude chat data
AI 2040: Plan A maps a coordinated US-China slowdown
Hassabis proposes a FINRA-style standards body for frontier AI
Jailbroken Gemini builds a working C2 server in six minutes
Grok Build CLI caught uploading entire repositories
Anthropic finds Claude's personality shifts across languages
Google releases Mantis Skills for agent security reviews
Microsoft warns AI means busier Patch Tuesdays
OpenAI's new flagship GPT-5.6 "Sol" reported deleting files unprompted
Researchers show frontier LLMs still readily jailbroken
Cursor zero-day disclosed after limited vendor response
"Context bombing" turns prompt injection into a defence
Boko Haram reported using frontier AI for propaganda
A fresh crop of security holes in AI coding agents
China's vulnerability database warns developers off Claude Code over alleged backdoor
Anthropic's GRAM offers an "off switch" for dual-use knowledge
Varonis "Rogue Agent" flaw exposes Google Dialogflow CX chatbots
Lilian Weng reframes recursive self-improvement around "harness engineering"
"HalluSquatting" turns AI coding assistants into botnet builders
IEEE Spectrum: reasoning models can be tricked into denial-of-service by "overthinking"
Lawsuit alleges xAI's Grok was used to generate child sexual abuse imagery
Anthropic interpretability team reports a "global workspace" inside Claude
GitHub's AI coding agent leaks private repositories when asked nicely
Discord's AI moderation bug wrongly banned thousands
Savi launches an app to counter realistic AI scams
Import AI: Fable writes the fastest GPU megakernel yet
Alibaba to bar Claude Code from 10 July amid Anthropic distillation dispute
SnortML brings machine learning into intrusion detection
AI voice clone costs a father his life savings
Proof of human: verifying real, unique people in the age of AI
Smooth-talking AI drives "first" end-to-end agentic ransomware attack
Startup sues Koi Security over AI-hallucinated espionage report
Devin Security Swarm scans whole codebases via Agentic MapReduce
Alibaba bans Claude Code in the workplace over alleged backdoor risks
Oxford lecture reframes AI as an agent, not a tool
AIEWF Day 3: autoresearch and the tension between AI and human agency
Two-thirds of enterprises hedged their AI strategy after the Fable 5 outage
Lightrun adds pull request production-impact assessment
Anthropic removes covert steganographic code from Claude Code
Red teamers turn Claude Desktop into a double agent
DeepSeek generates functional in-browser ransomware on request
Autoresearch: the feedback loop behind self-improving agents
Reliability comes from the system, not the agent
Researchers trick LLMs into harmful outputs via role model prompt injection
Mythos AI discovers Squidbleed: a decades-old memory leak in Squid proxy
Microsoft Copilot Autofix extends to Azure DevOps
Infosec professionals souring on autonomous AI pentesting
Claude Code is steganographically marking its requests
Morgan Stanley: making agents less autonomous cut reconciliation risk in half
Mozilla: indirect prompt injections pose severe risk to AI coding agents
Survey: IT incidents attributable to AI coding tools rising sharply
AI-generated code sparks a production confidence crisis
Survey: more AI code is running in production, despite persistent caveats
New attack shows why letting an AI browse for you is risky
Emily Bender sets the record straight on "Stochastic Parrots"
AI surfaces a wave of previously hidden security vulnerabilities
Amazon Q vulnerability: malicious Git repos can execute code and steal cloud credentials
Import AI 463: robot self-improvement through repeated practice
US lawmakers move to ban AI companies from selling health data
Attackers exploit SimpleHelp flaw to steal data from AI coding assistants
China's Z.ai claims GLM-5.2 matches Mythos on cybersecurity benchmarks
Runtime verification is the missing layer for production coding agents
Cloudflare Ships Open-Source Agent Skills Library for Zero Trust Deployment
Reward Hacking Benchmark: RL-Tuned Coding Agents Exploit Evaluation Flaws at 13.9%
Simon Willison: what happened after 2,000 people tried to hack his AI assistant
Hasp: Local-First Secret Broker with Time-Bounded Agent Access
Akrites: new coalition to protect open-source from AI-assisted supply chain attacks
Medical Diagnosis AIs Can Be Tricked into Revealing Whose Data Trained Them
Microsoft Uses AI to Link StealC and Amadey Malware Operations
Anthropic Mythos/NSA Tie to Project Glasswing Confirmed by Journalist
OpenAI Claims Cyber Model Outperforms Mythos on CyberGym Benchmark
Five Eyes Agencies Warn AI Will Escalate Cyber Incidents to Major Crises
OpenAI Launches GPT-5.5-Cyber and Daybreak Cyber Partner Programme
Nvidia Opens Humanoid Robotics Safety Lab
Claude Code Extended Thinking Output Is Encrypted and Not Verifiable as Authentic Reasoning
DeepMind Publishes AI Control Roadmap for Securing Internal Agents
Amazon VP Rejects Human-in-the-Loop as AI Governance Backstop
DiffusionGemma Transparency Audit Finds Non-Chronological Reasoning
Agent Hooks Enforce Guardrails Deterministically Inside Agent Loops
The Flat Curve Society Argues Against Frontier AI Deployment
Import AI 462: Research Shows AI Systems Achieving Superpersuasion
Import AI 462: Mapping Technical Paths to Artificial Superintelligence
AI Agent Uses Nuclear Strike in Wargame but Still Loses Overall
Anthropic export standoff shifts towards a resolution framework
GitLab 19.0 embeds agentic AI across the SDLC
Microsoft positions Windows to secure AI agents
Cloudflare Launches Temporary Accounts for AI Agents
G7 ends with no movement on the Anthropic export ban
Chainguard launches the Athena coalition to turn AI on open-source flaws
Nature Study Finds Early Evidence AI Use Degrades Human Cognitive Skills
Stack Overflow Blog: AI Agents Need Contract-Based Capability Declarations
Checkmarx SAST Engine Focuses on Post-LLM Code Quality
Anthropic export standoff hits day five as Mythos becomes the sticking point
MCP Gets Its Missing Enterprise Authorization Layer
OpenAI details Deployment Simulation for pre-release safety
Ars Technica: Advanced AI Hacking Capability Is Coming Regardless of Export Controls
Stack Overflow Blog: AI Agents Are a Confused Deputy with Access to Your Systems
Over 100 security experts sign a "Free Fable" letter as the export ban holds
KPMG's AI report becomes an accidental hallucination demo
Critical Microsoft Copilot "SearchLeak" flaw exposed 2FA codes and emails
Fable 5 standoff deepens as Anthropic flies staff to Washington
Researchers launch Sequent, betting alignment is not on track
Google DeepMind maps the road from AGI to superintelligence
Cornell finds 13 words can poison AI deep-research agents
Tenet's "Agentjacking" turns Sentry errors into code execution
NewCore raises $66M to give AI agents workplace identities
Amazon reportedly triggered the US shutdown of Fable 5 and Mythos 5
US government orders Anthropic to suspend Fable 5 and Mythos 5
Google sues Chinese cybercrime network that abused Gemini
Google's "faithful uncertainty" curbs confident hallucination
Survey finds AI-generated code is deepening a DevSecOps crisis
An AI agent bankrupted its operator while scanning DN42
Anthropic's Claude Fable 5 launch turns controversial
Chainguard's "greyware" scanner flags 52,000 packages
Chatbots argue harder, not better, when challenged
An AI agent runs amok in Fedora and elsewhere
A prompt-injection flaw in Claude Code risked CI/CD secrets
A €0.01 transfer could compromise a banking AI agent
Google Research proposes a framework for auditing machine unlearning
Lessons on AI trustworthiness from 500 years of journalism
Anthropic urges a global pause over self-improvement risk
