Industry
An X user says Amazon stops Grok Bot purchases
Amazon does not say that the stop is correct, but Musk says SpaceX will repay a user for errors.
BigGo FinanceClaimed, not confirmed
News of the day
15 October 2026
63 stories filed under this day.
Security
Perplexity makes its SPACE sandbox for AI agents open source
Perplexity reports that nine frontier models tried 118 times to break out of the sandbox and none did.
Agents
ChatGPT can view the Chrome tabs that a user has open
The user must tag Chrome in the ChatGPT web or desktop app first.
Security
Proof launched x401, an open protocol for AI agent authorization
If a term of the transaction is different from the signed term, the signature is not correct.
Industry
4 banks are part of a protocol that sets limits for personal AI agents
Customers select the agent, and the company sets limits for that agent.
Research
An MIT report finds only 33 percent of companies have board oversight
Companies with a cross-functional team put sustainability in decisions 57 percent of the time. With one function, the number is 13 percent.
Policy
Claude Haiku 4.5 sent an incorrect murder tip to Philadelphia police
Police did not read the tip until Anthropic told them on October 7.
Deals
Proofpoint agreed to buy Varonis for about $6.3 billion in cash
The companies plan to close the deal in the first half of 2027.
Models
Granite Switch 4.1 gives chat, decisions and voice from one endpoint
The 8B and 30B models use about the same GPU memory as the base model.
Security
OWASP moves excessive agency from sixth to third in 2026
Prompt injection stays first, and incident records give 25% of the weight.
Deals
New York City comptroller tells Anthropic to put safety before its IPO
Levine writes that Anthropic must first put AI safety and governance before its public share sale.
Research
At times, GLM-5.3 thinks about how its steps change training
The researchers did not cause this, and it occurred in tests that do not tell the model about training.
Apollo ResearchClaimed, not confirmed
Industry
Bank of America adds 1,000 contractors to repair bugs found by Mythos
The quantity of bugs is the problem, because a person must repair each one.
Industry
Anthropic gives 6 METR employees access to its systems
METR will tell all persons its results, and Anthropic cannot stop this.
Agents
GreenCore starts gtinfetch, a barcode check for AI agents
The company says that one call gives the result for 84 million barcodes.
Research
Jev selects its server and not 20 persons in 16 of 48 prompts
8 of 13 other models selected their servers in more of the 48 prompt versions than Jev did.
zachlandes.comClaimed, not confirmed
Models
Meta releases Muse Spark Open, a model with 816B parameters
A company must sign an agreement with Meta before it sells the model as a service if its revenue is more than $20 million in 12 months.
Models
Nscale adds the NVIDIA Nemotron 3.5 Super VL model
The open model has 120 billion parameters, and 12 billion of them are active.
Compute
The CoreWeave credit default swap spread is above 1,140 basis points
Gavan Nolan of S&P Global Market Intelligence says a level of 1,000 basis points or more is a sign of credit distress.
Devices
Signal president says Muse puts the data of contacts at risk
Meta says Muse uses public information and the data that the user gives to it.
Agents
An AWS agent exploited 33 of 37 new vulnerabilities in a test
AWS uses only new vulnerabilities, and the agent cannot get an exploit from the internet during the test.
Models
GitHub makes AI Scan for pull requests available
AI Scan finds vulnerabilities in code languages that CodeQL does not check, but its results cannot stop a merge.
Deals
Ramp raises $1.85 billion at a $60 billion value
Ramp says that its revenue for a year is more than $2 billion at this time, up from $1.5 billion in June.
Security
A check of 217 million web responses found 7,959 sites with AI headers
Most of these headers send agents to a different resource, but 30 sites sent injection text.
Deals
Three groups give $15 million to help mathematicians use AI
The Simons Foundation will give the money to projects after an open call for proposals.
Security
Descope releases Agentic Identity Hub 3.0 for internal AI agents
The release adds an approval step. Each agent has an owner and a clear task.
Agents
Flash Mixer gives AI agents an MCP server to mix Bitcoin
Elliptic found no other mixer with an MCP server.
Security
Researcher shows AI worm that moves through Copilot for Word
The worm changes financial figures and copies its payload into each new document.
Reverse Engineer Anything, Tracking AI-integrated Malware, Test-driving JevClaimed, not confirmed
Industry
Amazon stops the Muse agent of Meta and JPMorgan stops external agents
A website cannot know which user an agent works for or if that user let it make the payment.
Industry
A pilot added hundreds of millions of pages to the OpenAI index
The pilot used only about 1% of the network of Cloudflare.
Security
ServiceNow adds AI Passport to set limits for AI agents
In a demonstration, the agent sends an $18 refund but the system stops a $40 refund at runtime.
Security
The Aoki malware gets credentials of AI tools for code
A token that an attacker gets gives the identity and the API quota of the developer.
Deals
Cohere wants to get AISLE, which finds software vulnerabilities
In its first year, AISLE found more than 500 unknown vulnerabilities.
Research
A Snyk agent passes 90.2% of dependency upgrades in a new benchmark
Claude Code passed 59.6% of the same upgrades with no Snyk data, and 77.5% with it.
Security
GhostAction attack targeted more than 30,000 GitHub repositories
Subsequent waves on October 10 and 12 got no secrets, because GitHub stopped the workflow from October 9.
Security
Honeypot sensors record 1,423 tries to use a Langflow vulnerability
A user who has a login can execute commands on the server in versions before 1.9.0.
Policy
UK has no AI rules in its Cyber Security Bill at this time
Officials told MPs that they think about changes, but the committee chair told them that a bill must make law.
Security
A new benchmark tests if guardrails follow authorization
Many blocked operations come from check errors and not from correct denials.
arxiv.orgClaimed, not confirmed
Research
A new test shows that task checks find 0.09 of agent risks
A check of the workflow condition finds 0.96 of the same violations.
arxiv.orgClaimed, not confirmed
Security
Researchers test a guard that stops a reply only if it turns dangerous
Prompt labels and response labels agree on only 50.46% of items in a set of 41,944 pairs.
arxiv.orgClaimed, not confirmed
Research
A new test finds that AI agents do not work each time with emotional users
Data from the Tracer pipeline improves Qwen3 models at each size of model in the test.
Research
A study shows that accuracy is not a good test for AI agent teams
In the tests, accuracy showed that dangerous teams were safe, and that good teams were not safe.
arxiv.orgClaimed, not confirmed
Research
One change to a Claim Card changes 1 in 3 of the agent verdicts
The test used 900 runs of a GPT-5.4 agent on 60 claims with expert labels.
Security
Researchers jailbreak language models with a moral reframing attack
The success rate is high on the first try and increases with more tries.
arxiv.orgClaimed, not confirmed
Agents
A new benchmark finds that AI agents frequently do not release locks
Frontier models are not correct in more than 58% of closing steps.
Security
Censys finds 63% more AI software on the internet in 9 months
Industrial control systems on the internet show only 4% more in 2.5 years.
Agents
CyberArena tests AI agents in live attack and defense matches
The test has 40 challenges and 466 matches with 12 frontier agents.
Research
More documents cut RAG precision by 63% in a Wyoming system
Soft scoping by domain made the problem less, but hard partitioning was dangerous.
arxiv.orgClaimed, not confirmed
Research
One warning sentence helps AI agent groups with incorrect context
The warning recovers a maximum of 0.18 accuracy with voting.
Research
Small AI models suit short agent tasks, and large models long tasks
The researchers measure energy, cost and time for each completed task, not for each call.
Research
Study of 2,500 agent runs finds 20 failure modes in agent teams
In the study, agents do not check the facts of other agents unless a person tells them to.
arxiv.orgClaimed, not confirmed
Research
One closed model gets consensus with no messages and 6 open models do not
A network with tight clusters gives less consensus for all models.
arxiv.orgClaimed, not confirmed
Research
2 policies in a game find where preference models give incorrect results
The system uses disagreement in a group of preference models as a signal of error.
arxiv.orgClaimed, not confirmed
Research
A 240-task benchmark shows that agent increases stay on one file type
The top configuration got 42.25 in all and 30.62 on tasks with more than one file type.
arxiv.orgClaimed, not confirmed
Research
A food safety test shows a mean attack success rate of 25.51% in 12 LLMs
A test with one jailbreak method can underestimate the risk, as the best methods reached 60.59% on average.
arxiv.orgClaimed, not confirmed
Research
A verifier reward increases code proof verification from 66.1% to 93.5%
An audit rejects each result that changes the specification.
arxiv.orgClaimed, not confirmed
Research
Activation memory used stored facts in at most 15% of tests
A stored dislike can also come back as a like, unless the writer subtracts the opposite.
arxiv.orgClaimed, not confirmed
Research
An attacker can wait for certification and then start an attack
Researchers show that a finite test cannot find an attacker that waits for certification.
arxiv.orgClaimed, not confirmed
Research
Coding agents can copy some biological AI tools with no internet access
The agents could copy one very complex tool only to a small degree.
arxiv.orgClaimed, not confirmed
Research
Neural predictors do not give more accurate gene-pair results than a sum
Large differences from the sum show large errors.
arxiv.orgClaimed, not confirmed
Research
The CASI method gives a 3.51% increase in a 7-day test on 10% of traffic
CASI gives credit to each retrieval step when the system gives one reward for the full set.
arxiv.orgClaimed, not confirmed
Research
The model with the top score on questionnaire items gets 0.813, level with a baseline
A Gaussian copula made from other respondents gets 0.824.
arxiv.orgClaimed, not confirmed
Research
Training one trait in a language model can change other traits
No one predictor was the best, and the researchers say that generalization remains hard to predict.
arxiv.orgClaimed, not confirmed