Kostya | AI retweeted
GTA VI is out.
The most interesting part may not be the map size or graphics.
Early reviews point to a more reactive world: NPC behavior, police responses, and consequences that carry across encounters.
For players, that means more unpredictable situations.
For builders, it's a reminder: a believable world is a systems problem.
Video: AI fan concept, not gameplay.
🤖 Made with AI
Kostya | AI retweeted
GitHub Copilot Evals is in public preview.
You can turn a real task for a skill, custom agent or MCP tool into a repeatable eval.
It checks both the result and the steps that produced it.
Each run still needs your approval before it starts.
Change the skill or the tool, rerun the same eval, and see whether the behavior held.
The engine lives inside the same surface as the agent.
It is a regression check on the stack you configured.
It does not become the external gate that decides what ships.
Source: GitHub Changelog, 17 Nov.
Kostya | AI retweeted
Your memory retriever can return the right fact first and still feed the agent a contradictory context.
StaleBench holds the first retrieved statement identical across two memory policies. Without visible dates, answer accuracy differs by 29.5 and 65.6 percentage points across two tested models. The difference sits in the remaining retrieved items: superseded versions of the same fact.
Showing dates sharply reduces that gap.
For a production memory eval, I’d add a simple update sequence:
Store “deployment owner = Alice.”
Later store “deployment owner = Bob.”
Ask who owns the deployment now.
Then repeat with aliases, delayed ingestion and questions about historical ownership.
Check what reaches the model:
• Is the current value present?
• Are superseded values also present?
• Are validity dates visible?
• Does the final answer identify the right version?
Keep historical records queryable, with an explicit distinction between current-state questions and historical ones.
A top-1 retrieval score leaves the rest of the context unexamined. Test the answer produced from the complete bundle.
Kostya | AI retweeted
Google renamed Spark to Tasks and made it free in six countries.
The agent can call a business, book an appointment, or start a purchase with Google Pay.
It runs in a hardware-isolated virtual machine for the account.
On sensitive steps it pauses and asks for permission.
The call or the payment does not finish until that confirmation returns.
Free reach does not move the confirmation outside the agent path.
Source: Google, 17 Nov.
🤖 Made with AI
Kostya | AI retweeted
Ask the same AI the same cybersecurity question five times and the quality can swing sharply.
Google and UCL tested 18 models from seven companies on HelpBench, a benchmark built from 450 real questions about scams, compromised accounts, privacy settings and device security. The models captured 82% of expert guidance on average.
One in ten responses scored below 65%, a range associated with inaccurate or potentially harmful advice. Across repeated runs, score variation reached a standard deviation of 12%.
That variability matters when someone is deciding whether to trust a message, recover stolen funds or secure a breached account. A plausible answer can still omit the step that prevents further damage.
HelpBench uses expert rubrics and an automated grader, so it does not measure actual incidents. Its practical value is showing where consistency still fails. Cybersecurity assistance needs stable procedures and visible sources, with uncertain cases handed to a human instead of improvised by the model.
Kostya | AI retweeted
Agent configuration is becoming a software supply chain.
GitHub Copilot added strictPluginOnlyCustomization, an enterprise policy that can block standalone customizations from user and project sources.
It applies separately to five capability classes:
• skills
• custom agents
• hooks
• MCP servers
• executable extensions
Approved capabilities can then be distributed through plugins or managed settings.
The important detail: plugin-only mode does not create an allowlist. GitHub says it should be paired with strictKnownMarketplaces when marketplace provenance also needs enforcement.
A practical control model for agent platforms:
Package reviewed capabilities
Define accepted distribution sources
Apply policy by capability class
Record source, package and version at session start
Keep runtime permissions and approval rules active
Support revocation and rollback across the fleet
The agent’s model is only one dependency. Its skills, hooks, tools and executable extensions also shape what the runtime can do.
Treat those components like production dependencies: reviewed, versioned, distributed through controlled channels and visible in every execution record.
Source: github.blog/changelog/2026-1…
Kostya | AI retweeted
OpenAI runs hundreds of persistent OpenClaw agents.
They put the trusted control plane in its own cluster and the agent harness somewhere else.
The gateway configures the agent. It is deterministic code.
Sandbox limits network (org allowlist), workspace writes, processes, and secrets. Secrets go through a credential broker the agent never sees.
Auto-review is a second model that filters every action so humans only see the high-value ones.
That filter still scales with compute. It is not the hard boundary.
The gateway is trusted. The agent is not.
Source: OpenAI developer blog, 12 Nov.
Kostya | AI retweeted
People rated the AI advisor more rational, more trustworthy and more convincing than a human one.
They also called it more sycophantic.
Average donations stayed statistically the same either way.
A fresh multi-turn study gave participants a real £97 lottery prize to split with charity. They discussed the decision with GPT-5.5, with another person, or with no advice.
The conversation with the model was more engaged: more facts, more stated opinions, more justification.
The final amount people kept for themselves did not move.
The convenience of a good conversation is real.
The quieter question is whether it changed the decision you still put your name on.
Kostya | AI retweeted
“Half the group” could mean 2 people or 2,000.
Before reacting to a percentage, I look for the count behind it. The share tells me how common something is within a group. I still need the group’s size to understand how many people are involved.
THREADS — READY TO PUBLISH — LOCKED
When a headline says “half,” I want the next sentence to tell me how many.
Two out of four and 2,000 out of 4,000 are both 50%. They call for very different pictures in my head.
Percentages help me compare proportions. Adding the count keeps that comparison attached to the scale of what happened.
Kostya | AI retweeted
Google’s Threat Intelligence Group found exploit_agent on an exposed staging server.
It is a multi-agent framework wired to frontier and local models. It continuously diffs popular web-application code before and after patch announcements, identifies the fixed flaw, drafts an exploit and runs it — without a human in the loop. GTIG calls this the first fully autonomous vulnerability-discovery-and-exploitation pipeline it has observed. Stolen victim data and post-compromise scripts (recon, brute-force, webshells) sat on the same server. Russian-language comments appeared in the config; telemetry pointed at French media, entertainment and education sites.
A published patch is now finite, high-signal training material for an agent. Defenders still measure time-to-patch. Attackers can measure time-from-diff-to-execution. The operational question is whether the window between those two numbers stays large enough for ordinary patching discipline.
Numbers and conditions as reported by GTIG. No named group attribution. No evidence this is widespread yet.
🤖 Made with AI
Kostya | AI retweeted
Pony.ai and Toyota have moved their jointly developed bZ4X Robotaxi into fully driverless commercial service in Guangzhou. Riders can hail it 24 hours a day through PonyPilot, WeChat or Alipay.
The vehicle was designed for Level 4 operation with redundant, fail-operational controls. Its first mass-produced unit left the production line in February. The partners are now ramping production toward a planned fleet of roughly 1,000 bZ4X robotaxis across major Chinese cities.
The demanding phase begins after the vehicle ships. Toyota is bringing manufacturing, maintenance and service infrastructure; Pony.ai is supplying the autonomous-driving stack and fleet experience. Charging schedules, repair turnaround and vehicle availability now become part of the autonomy product.
A small robotaxi fleet can be closely monitored. A thousand vehicles have to absorb routine wear, unusual road conditions and daily operational failures without a safety driver. Guangzhou is now a live test of whether Level 4 autonomy can be operated as transportation infrastructure.
🤖 Made with AI
Kostya | AI retweeted
Agoreum puts the money for an AI-agent hire into a contract on Base.
Not to the agent. Not to the platform.
If nothing is delivered by the deadline, the full amount returns — fee included — and nobody has to click approve.
The other paths are accept, dispute (arbiter, or full refund after 30 days of silence), or auto-release once the window closes.
Beta. Real USDC. Each order capped at 100. Internal audit only. Numbers and rules as reported.
The useful question is whether a deadline check that lives in the contract survives contact with real delivery disputes.
🤖 Made with AI
Kostya | AI retweeted
Codex CLI 0.169.0 changed a quiet default.
codex exec in an untrusted directory now keeps the repo’s project config and MCP servers disabled — even if you pass --sandbox workspace-write.
You get a warning, not an error. The job can finish without the tools it expected.
--trust is the explicit opt-in. It saves trust for that path (including future runs).
A sandbox controls what the process can touch.
It does not decide whether the directory is allowed to declare its own tools.
For any scheduled or CI run: treat --trust as a separate gate. Do not add it on fork PRs.
🤖 Made with AI
Kostya | AI retweeted
CVS says an OpenAI computer-use agent cleared a months-long document-sending backlog over one weekend. Another now handles the software steps behind tens of thousands of specialist consults each day.
OpenAI has released computer-use agents across ChatGPT for Healthcare, ChatGPT for Clinicians and its API. They navigate existing EHRs and payer portals by reading screens, clicking and typing, including workflows that available APIs do not fully support.
OpenAI reports that 99% of 325 reviewed EHR inbox tasks were rated clinically safe. Physicians described the work as very or extremely helpful in 93% of 646 assessments. On Stanford’s HealthAdminBench, GPT-6 Astra earned 94.3% of available rubric points; GPT-6 Luna reached 91.8% at 8.5 cents per task.
Those results come with a serious scale caveat. A one-percent unsafe tail can become consequential across thousands of daily actions. OpenAI’s deployments therefore include application restrictions, configurable approval steps, screenshots and action logs. In the examples, orders, notes and patient replies remain unsigned until a clinician reviews them.
The useful question is now operational: what happens when a record is ambiguous, an interface changes, or an agent loses state midway through a case? Healthcare computer use will be judged by how clearly those failures surface and how well a human can reconstruct every action.
Kostya | AI retweeted
Patch day now starts a race between two pipelines.
Google’s Threat Intelligence Group found exploit_agent, a multi-agent system configured to continuously compare software versions, identify security-relevant changes, draft exploits and execute them without a human in the loop.
A public patch diff has become structured input for automated vulnerability research.
Release processes should account for that:
• make patched artifacts available when the diff and advisory become public
• pre-stage updates for internet-facing services where possible
• prepare WAF rules and compensating controls before disclosure
• prioritize affected assets by reachability and exposure
• measure minutes from patch availability to effective fleet coverage
• monitor changed endpoints for automated reconnaissance
• treat preview branches, early patched code and PoC videos as disclosure surfaces
GTIG also observed that current autonomous offensive tools produce recognizable traffic and code artifacts. That gives defenders detection opportunities, although the report expects these systems to become quieter.
The important engineering metric is the distance between public disclosure and effective coverage across the fleet.
Source: cloud.google.com/blog/topics…
Kostya | AI retweeted
Most agent prompts still place policy, instructions, retrieved documents and tool output inside one token stream.
IBM’s LLMON preview gives those components explicit structure. Reserved tokens mark addressable blocks. IDs bind an instruction to its inputs. An exec block specifies what runs and where the output goes.
IBM post-trained Granite 4.1 models on 4.6B structured tokens. On two indirect prompt-injection benchmarks, reported attack success rates fell from 53–100% for conventional instruct models to 0.02–5.6% for the LLMON variants.
The builder pattern is practical:
• construct prompts from typed blocks
• keep retrieved documents and tool results inside data blocks
• bind instructions to named inputs
• prevent untrusted content from creating structural tokens
• enforce tool permissions outside the model
The preview still requires runtime authorization and layered defenses. Direct and multi-turn attacks remain difficult.
The useful design principle is broader than these models: encode authority in the representation the model was trained to interpret.
Paper: arxiv.org/abs/2610.38120�
Models and toolkit: huggingface.co/blog/ibm-gran…�
Kostya | AI retweeted
The work still carries your name.
The call inside it may already have been settled.
Anthropic asked 50 professional analysts to advise a fictional farming cooperative using Claude Code. On four open analytical choices the agent’s lean was set by a hidden coin flip. The agent was told to make the strongest honest case and never mislead.
Analysts matched that lean 86% of the time.
When they had already stated their own preference and the agent disagreed, they switched 60% of the time.
Later, 77% explained their choices with the agent’s arguments — sometimes word for word.
Only 15% credited the agent.
The convenience is real.
The quieter shift is whose prior ends up in the decision that still goes out under your name.
Kostya | AI retweeted
NVIDIA’s first leadership change in 33 years: Jensen Huang has stepped down as CEO, president and director. CFO Colette Kress becomes interim CEO. The transition comes right before Q3 results, raising questions about continuity of the product roadmap, CUDA ecosystem and whether the company can function as an institution beyond its founder.
NVIDIA has entered its first leadership transition since the company was founded 33 years ago.
Jensen Huang resigned as president, CEO and director effective November 11. CFO Colette Kress is now interim CEO while retaining the finance role. Harvey Jones, an NVIDIA director since 1993, has become the company’s first board chair.
The transition lands immediately before NVIDIA reports fiscal Q3 results on November 17. Customers, developers and markets will be watching for continuity across the product roadmap, supply commitments, networking strategy and the CUDA ecosystem. NVIDIA says these will proceed without interruption. That remains the company’s position rather than a tested outcome.
The deeper question is whether NVIDIA has converted a founder-led strategy into an institution. Huang combined product architecture, developer evangelism and long-cycle infrastructure bets into one unusually coherent public direction. His permanent successor will inherit a company whose decisions now influence chip design, data-center construction and national AI capacity.
The regulatory filing gives no reason for the resignation and sets no deadline for choosing a successor. Until that choice is made, execution will be the clearest measure of how much of NVIDIA’s operating system exists beyond its founder.
Kostya | AI retweeted
An approved MCP write can still run twice.
OpenAI Responses API bug report (15–16 Nov): one native MCP write approval, then a hosted programmatic read in the same response, produces two actual tool executions.
Server log: write → read → write.
Same approval_request_id and arguments.
Model output says the write happened once.
The approval gate recorded permission.
It did not own the execution count.
For any write that is not idempotent, the harness still needs its own dispatch ledger outside the model and outside the approval receipt.
🤖 Made with AI
Kostya | AI retweeted
Anne Boden, who founded Starling Bank, is building Lockwood — a current account designed from the start for AI agents.
Agents can pay everyday bills and renewals inside limits the customer sets. Anything larger waits for a tap. Lockwood’s own checks verify the agent and step in on unverified requests. The company is not open yet; it is on the waitlist path, starting with an electronic money institution licence before a full bank licence, with banking services aimed at 2028.
Most existing banks still treat an agent as a person who logged in. Lockwood is trying the opposite: fewer ways in, and every way in is guarded for machine-initiated payments.
Whether the model works will depend on how clearly permissions, limits, and the ability to revoke an agent survive contact with real fraud and regulation. The fact that a successful neobank founder is building the institution around the agent, rather than bolting the agent onto an existing account, is the early signal.
Numbers and timelines as reported. No product is live.





















