Models
Grok Imagine Video 1.5 Lite has a price of $0.14 for each second
Artificial Analysis writes that it makes a 10 second 1080p video in a median time of 60.5 seconds, the lowest time at its quality.
News of the day
8 October 2026
40 stories filed under this day.
Security
Anthropic started free AI scans for open-source vulnerabilities
A model writes each report and no person examines it, and Anthropic writes that some reports can be incorrect.
Industry
114 lawmakers tell Google to stop a Spirit Airlines data deal
Google wants to pay $10 million for about 100 million emails, but the lawmakers write that removal of names does not always make data anonymous.
Security
Microsoft makes MXC available to limit the files an AI agent can use
A developer can select the Locked Down level to stop an agent from the use of the internet and files.
Research
Vitalik Buterin agrees with a plan to move crypto funds to new addresses
Justin Drake writes that AI mathematics could break wallet signatures, but Buterin writes that it is not necessary to move funds today.
cointelegraph.comClaimed, not confirmed
Industry
GPT-6 Sol Daybreak Blue is first on the Cyber Index
One task is $1.77 for this model and $11.67 for Grok 4.7 xhigh.
Agents
Anthropic released Claude Dashboards in beta
A user can click a number to see the query that made it, and can move the dashboard to tools such as Grafana or Hex.
Industry
Jefferies found that one third of sites stop the Meta Muse agent
Jefferies writes that access to websites is the primary problem, and it can slow the revenue that Meta wants.
Benchmarks
Grok 4.7 is first in the new legal AI test with 9.4%
Artificial Analysis says a hallucination is material if it can mislead a reader, for example with an incorrect contract date.
Security
Incognia started a fraud check for AI agents
Incognia writes that a legitimate agent does not show that the customer wants the task.
Security
A worm was found in Tensorlake SDK version 0.5.144
Researchers say a user must disable the token monitor of the worm before the user revokes tokens.
Models
Nano Banana 2.1 from Google has a price of $0.0336 for each 1K image
The price at 1K resolution is half the price of Nano Banana 2.
Security
Goodfire says its monitor for AI agents costs $51 for 1,500 sessions
In tests, the monitor found 94% of malicious hacking sessions but sent 8.7% of safe sessions for a check.
Agents
The US government started a new AI website for its services
The White House has a plan to let users do tasks, such as passport renewal, with AI agents before the end of 2026.
Industry
Atomic Canyon made NIVA, an AI tool for all 94 US reactors
Most of the 94 US reactors got an AI tool for their work.
Security
A new Claude Code session sent data out with no confirmation
The confirmation is on one path, and the attack used a different path to the same result.
straiker.aiClaimed, not confirmed
Benchmarks
AI agents wrote GPU code with 1.57× the speed of `torch.compile`
The author tells that the correct test was more work than the fast code.
towardsdatascience.comClaimed, not confirmed
Security
Cogent made a tool to find attack paths to the data of a company
Cogent tells that its study found 3 new attack paths for AI agents for each 1 for human attackers.
Agents
Crossmint made one toolkit for payments by AI agents
Crossmint tells that the toolkit has 1 API for the work of 4 or more vendors.
Security
An attacker got an ASOS employee login and then customer data
ASOS tells that the attacker did not get payment information.
Security
One chat message gave Zenity control of all AgentCore agents
AWS made changes for part of the problem, and Zenity tells companies to give each agent a role with only the access it must have.
the-decoder.comClaimed, not confirmed
Policy
Ten AI developers agreed to change how they use personal data
The ICO also wrote to OpenAI, Anthropic and Meta about AI agents that did not obey their safeguards.
Security
The first AWS patch for AgentCore agents did not remove the risk
AWS corrected the default permissions of the agents between June 22 and September 29, Zenity wrote.
www.csoonline.comClaimed, not confirmed
Security
Cycode found a vulnerability that let an MCP server steal OAuth credentials
Upgrade the SDK to version 2.2.0 or 1.30.0 to remove the vulnerability.
Security
Talos found 84 malware samples that put text in code for AI tools
The text works in only about 35% of tests at most, and it is always plaintext, which tools can find.
Research
A local model is equal to hosted models for 4 of 5 call sites
The authors tell that 2 hosted call sites give the same live result for 28% of the spend.
arxiv.orgClaimed, not confirmed
Security
President Lee thinks AI agents were used in attacks on 7 or more firms
The attacks made the data of 68,000 or more customers open to the attackers.
The SOFX ReportClaimed, not confirmed
Agents
AgentMemGate keeps plans out of the memory of AI assistants
The authors tell that, with the gate, 95% of tasks are correct, not 65%.
arxiv.orgClaimed, not confirmed
Security
A rule monitor found about 70% of attacks on AI agents in a test
The authors tell that the monitor also gave an alarm for 29.3% of safe runs.
arxiv.orgClaimed, not confirmed
Security
An attack on world models can give robots dangerous policies
The authors put malicious prompts in data that looks safe, and the attack works only after a world model uses the data.
arxiv.orgClaimed, not confirmed
Security
A guardrails proxy increases injection recall from 25.0% to 94.5%
The authors found a false-positive rate of 0.1%, and the system limits the model to a closed set of commands.
arxiv.orgClaimed, not confirmed
Models
Cloudflare releases Clef, open models that select from alternatives
In tests by Cloudflare, the 9B Clef-Flash model has a median latency of 38.8 ms, against 209.3 ms for the 27B Clef model.
Research
AI agents and humans get the same average, with different decisions
The authors found that humans, DeepSeek and Qwen3-Max all cut standard delivery by 48 to 53 percentage points. The results are different for each value of time.
Bioengineer.orgClaimed, not confirmed
Research
Only 0.7% of agent episodes in a hard test have no safety event
The authors found 12.3 times more collisions in real-time tests than in static tests. The agents completed 91% to 94% of the tasks.
arxiv.orgClaimed, not confirmed
Research
A test with 222 tasks shows that agents do not keep to a set time
The authors found a typical error of 2.9 times for Fable 5.1 in Claude Code, but only 1.2 times for GPT-6 Astra in Codex.
arxiv.orgClaimed, not confirmed
Security
A new attack agent goes around the guardrails of 4 AI models
The researchers find a success rate of more than 95% for the attack.
arxiv.orgClaimed, not confirmed
Research
Agents in a multi-agent system start to have the same behavior
2 easy methods to add diversity did not decrease the risk.
arxiv.orgClaimed, not confirmed
Research
An AI agent gives high confidence on 56% of high-risk tasks
The researchers find that this reporting destroys 68% of the gains from delegation.
arxiv.orgClaimed, not confirmed
Research
Reasoning AI agents do not select at random with the same inputs
An instruction to randomize independently decreased the correlation between agents with the same input, but did not remove it.
arxiv.orgClaimed, not confirmed
Security
The WebMirage attack hijacks web agents in 91.9% of tests
The attack can also get around 3 agent-level defenses.
arxiv.orgClaimed, not confirmed