Futureweb

Security

AWS's 89 percent is stage S3 on a benchmark it chose: 819 of 920 tasks. The same table's S4 is 37.8 percent. General availability on October 22 still asks a person to approve the patch.

AWS's 89 percent is 819 of 920 benchmark tasks at the stage that keeps tests. The stage that also fixes the chosen bug is 37.8 percent. People still approve.

The short version

On the October 5 security post, 89.0 percent is stage S3 of CyberGym-E2E: 819 of 920 tasks inside 90 minutes, where the patch still passes the project's tests. Stage S4, also fixing the historical bug the benchmark selected, is 37.8 percent. The October 22 GA post says every remediation starts with a person approving it. This page does not describe an exploit.

89 percent is one cell in a table

On October 5, 2026 Alexander Greaves-Tunnell and Neha Rungta posted that AWS Continuum for code vulnerabilities had passed 819 of 920 tasks on CyberGym-E2E within 90 minutes, which they call an 89.0 percent end-to-end success rate, 23.1 points above a previous public high of 65.9 percent. The same post's table splits the benchmark into stages. S3, preserving tested functionality, is the 89.0 percent. S4, also repairing the historical vulnerability the benchmark selected, is 37.8 percent, against a previous public high the post states as 26.2 percent. The New Stack's headline, that AWS says the agent fixes 89 percent of bugs but there is a catch, is a fair compression of that gap. The catch is not a critic's invention. It is the next row.the Brotli identifier is a different advisory.

What the benchmark asks, without the method

AWS says each task puts an agent in a container with a vulnerable revision of a real project and does not hand it a description, a proof of concept, a crash log, or the original patch. Within 90 minutes the agent is supposed to submit an input that demonstrates a vulnerability and a source patch. The benchmark then scores that submission in stages: a crash, a patch that stops that crash, tests that still pass, and the historical bug. This page stops at those stage names. It does not describe an input, a crash, or a patch. A success rate is a count of tasks. It is not a walkthrough.the advisory form is where a severity belongs.

The other percentages are AWS's too

The post's table gives S1, finding and reproducing a vulnerability, as 92.5 percent against a previous 67.9. S2, repairing the crash the agent itself produced, as 89.6 against 66.2. S3 as 89.0 against 65.9. S4 as 37.8 against 26.2. It also says that without the 90-minute cap the end-to-end rate reached 93.7 percent. Every one of those figures is AWS reporting its own run on a benchmark it says it chose. This page did not rerun the 920 tasks. A vendor's leaderboard is a vendor's leaderboard even when the arithmetic is printed in full.a CVE identifier is not this percentage.

October 22 is the product, October 5 is the number

The general-availability post, by Esra Kayabali on October 22, says Continuum finds and validates code vulnerabilities and delivers fixes as patches a team can automate or as pull requests. It says the service confirms which findings are exploitable in the customer's environment, ranks them, and proposes a fix. It also says every remediation starts in a mode where a person approves the change before it ships. A later mode, in which the customer might let some classes apply themselves, is written as future. The 89 percent is not in that post as a service-level agreement. It is a benchmark result from October 5, seventeen days before GA.Node's October release is a version list.

What the headline collapses

Fixes 89 percent of bugs is not what the table says. The table says 89 percent of these 920 tasks reached S3, and 37.8 percent reached S4. S3 can pass while the historical bug the benchmark cared about remains, which is why S4 exists. A reader who buys the headline and skips the row will think the agent closed the bug the dataset was built around. AWS's own numbers say that happened on a minority of the tasks. The New Stack putting a catch in the headline is the useful part of the wire item.

Who is in scope

AWS customers of the GA service, who on the October 22 post still approve a pull request. Readers of the October 5 post, who are looking at a benchmark rather than a contract. People who do not run the service are not patched by the percentage. A public benchmark score does not attach to a repository the customer has not connected.

What this page will not add

No reproduction of a task. No statement that 89 percent of all open-source bugs are fixed. No claim that the future unattended mode is on. The two AWS posts and The New Stack are enough. The October 5 post is the table. The October 22 post is the approval rule. The New Stack is the headline that refuses to let 89 percent travel alone.PHP's later note is a different bug.

A second AWS percentage is a different test

A separate October 15 post about capturing flags on newly disclosed vulnerabilities is not this table. This brief does not fold that 33-of-37 figure into the 819-of-920 figure. Different tasks, different post, different day. One product can have two benchmarks. They are not one percentage. Asked directly, Does 89 percent mean Continuum fixes 89 percent of bugs in the wild? No. It is 819 of 920 tasks on CyberGym-E2E, at stage S3, under a 90-minute limit, in a benchmark AWS says it chose. The New Stack's headline flags the catch. The catch is in AWS's own table.

The limit

S3 is 819 of 920. S4 is 37.8 percent. GA asks a person to approve. The record is those three sentences. The short answer this page will stand on is this: On the October 5 security post, 89.0 percent is stage S3 of CyberGym-E2E: 819 of 920 tasks inside 90 minutes, where the patch still passes the project's tests. Stage S4, also fixing the historical bug the benchmark selected, is 37.8 percent. The October 22 GA post says every remediation starts with a person approving it. This page does not describe an exploit.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. AWS Security Blog
  2. AWS News Blog
  3. The New Stack

Questions

Does 89 percent mean Continuum fixes 89 percent of bugs in the wild?

No. It is 819 of 920 tasks on CyberGym-E2E, at stage S3, under a 90-minute limit, in a benchmark AWS says it chose. The New Stack's headline flags the catch. The catch is in AWS's own table.

What is stage S4?

AWS says S4 checks whether the patch also fixes the specific historical vulnerability the benchmark selected. Continuum's figure there is 37.8 percent, against a previous public high AWS states as 26.2 percent.

When did the service become generally available?

The AWS news post is October 22, 2026. The 89 percent post is October 5. The New Stack's October piece is an account of the percentage, not the GA clock.

Does the GA post let the agent merge its own fix?

It says every remediation starts in a mode where Continuum proposes the fix and a person approves it. A future mode is described as future. It is not the mode the post says is on.

Does this page explain how the agent demonstrates a flaw?

No. The benchmark's definition includes a demonstration and a patch. The rates are the result AWS published. The demonstration stays in the benchmark.